Research ArticleContext-Aware Cloud Service Selection Model for Mobile CloudComputing Environments

Xu Wu 12

1School of Computer Electronics and Information Guangxi University Nanning China2Key Laboratory of Multimedia Communication and Network Technology Guangxi University Nanning China

Correspondence should be addressed to Xu Wu xrdz2006163com

Received 29 August 2017 Revised 3 January 2018 Accepted 21 January 2018 Published 5 March 2018

Academic Editor B B Gupta

Copyright copy 2018 Xu Wu This is an open access article distributed under the Creative Commons Attribution License whichpermits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Mobile cloud computing (MCC) has attracted extensive attention in recent years With the prevalence of MCC how to selecttrustworthy and high quality mobile cloud services becomes one of the most urgent problems Therefore this paper focuses onthe trustworthy service selection and recommendation in mobile cloud computing environments We propose a novel serviceselection and recommendation model (SSRM) where user similarity is calculated based on user context information and interestIn addition the relational degree among services is calculated based onPropFlow algorithmandweutilize it to improve the accuracyof ranking results SSRM supports a personalized and trusted selection of cloud services through taking into account mobile userrsquostrust expectation Simulation experiments are conducted on ns3 simulator to study the prediction performance of SSRM comparedwith other two traditional approaches The experimental results show the effectiveness of SSRM

1 Introduction

Mobile cloud computing (MCC) is a new computing modelwhere cloud computing (CC) is integrated into mobilecomputing environments The new computing model breaksthrough the resource limitation of mobile terminals bymoving data processing and storage from mobile device tocloud service platforms via wireless networks Rich mobileapplications can be easily created and accessed just based onweb browser on the mobile devices [1] The architecture ofMCC is shown as in Figure 1 Internet content providers inFigure 1 put video information games and news resourcesin appropriate data centers in order to provide users withmore rich and efficient content services Mobile users usethe wireless connections to access the data centers of publicclouds over the Internet Data centers of public clouds are dis-tributed in different locations and provide users with elasticand scalable computing or storage services In addition somemobile users with the demand for higher privacy protectionlower network latency and energy consumption can connectto cloudlets via local area networks A cloudlet located at theedge of the Internet is a mobility-enhanced small-scale cloud

datacenter which extends cloud computing infrastructure[2]

MCC is derived from cloud computing thus it inher-its the advantages of cloud computing such as dynamicdevelopment of mobile applications resource scalabilitymultiuser sharing and multiservice integration But quiteapart from that it has also some problems including resourcelimits weak battery strength user mobility and low networkcoverage With the prevalence of MCC more and moreInternet users from mobile devices use cloud computingservices through wireless interface (eg GPRS3GWiFi) Asshown in Figure 1 there exit abundant functionally similarcloud services provided by different cloud service providersTherefore how to select appropriate cloud services formobileusers becomes one of the most urgent problems

To address the problem of cloud service selection and rec-ommendation researchers have done many notable works incloud computing literature [3ndash7] The most of existing meth-ods mainly depend on ranking Quality-of-Service (QoS) toselect an optimal cloud service from a set of functionallyequivalent cloud services Quality-of-Service (QoS) usuallydescribes nonfunctional performance of cloud services QoS

HindawiWireless Communications and Mobile ComputingVolume 2018 Article ID 3105278 14 pageshttpsdoiorg10115520183105278

2 Wireless Communications and Mobile Computing

Mobileusers

Accesspoint

Networkoperator

Internet serviceprovider

Internet

Data center

Public cloud

Internet contentprovider

Cloudlet

Wirelessconnections

Figure 1 Architecture of mobile cloud computing

values of cloud services provide valuable information to assistdecision-making [3]

The challenge here is to recommend an optimal mobilecloud service based on dynamic QoS properties In real-world applications QoS inMCC is referring to dynamic QoSproperties (packet loss ratio end-to-end throughput delayetc) which are affected by the user context information [8]Context information includes location time resource ability(processing power memory or battery capacity) bandwidthand status (online or offline) For example limited bandwidthand resource ability can make the values of QoS proper-ties degrade Naturally consumers hope to select the mosttrustworthy cloud service among abundant candidates byconsidering the dynamic context information over multipletime periods However the works done for cloud computingrarely focus on investigating the influence of user contextinformation on service selection User mobility in MCC canlead to the dynamic changes of user context informationTheselection results of mobile user may vary according to thechanges of user context information Therefore user contextinformation plays an important role when designing suchselection and recommendation algorithms

Moreover the existingmethods always assume that cloudservices are independent and ignore the fact that the relationamong cloud services has important effect on the accuracy ofrecommendation resultsThe related serviceswill be probablyselected by the same user based on the intuition [9] Forexample in tourism service scenario the users purchasingairplane ticket service probably prefer to purchase car rentalservice in the future

To attack this critical challenge the algorithm for MCCis expected not only to satisfy the required service level byuser but also to adapt to dynamic changes of user context

information In the paper wemake use of user-based collabo-rative filtering method to propose a novel context-awareservice selection and recommendationmodel (SSRM) whichcombines the user context information user interest histori-cal service usage experiences (QoS value or customer rating)of cloud services and the relation among cloud services torank the cloud services Collaborative filtering technologies[10] in the field of online recommendation system offer us astrong theoretical foundation to deal with the service selec-tion and recommendation problem Recently they have beensuccessfully applied to predict themissingQoS value in cloudcomputing Collaborative filtering technologies are catego-rized as user-based collaborative filtering approaches item-based collaborative filtering approaches and their fusionapproaches In addition SSRM only selects trusted mobilecloud service as candidates for mobile users Different usershave different trust requirement the personalized serviceselection and recommendation method is thus required bydifferent users [3] To support personalized selection ofcloud services SSRM takes into account mobile userrsquos trustexpectation through using a trust threshold

Within the SSRM there are a fewmodules Figure 2 showsthe system architecture of SSRM which supports personal-ized selection and recommendation of cloud services

A user requesting service selection and recommendationis called target user in the model A user providing historicalservice usage experiences is called training user Our aim isto get the service ranking results from SSRM by predictingthe missing service usage experiences for target user SSRMinvolves the following key steps

Step 1 A target user requests service selection and recom-mendation

Wireless Communications and Mobile Computing 3

User similaritycomputation

Find similar user

Initial rankingresults

Relational gradecomputation

Ultimate rankingresults

User 1

User 2

User m

Mobile cloud netw

orking infrastructure

Trust filter

Context similarity

Interest similarity

Figure 2 Architecture of SSRM

Step 2 User similarities between target user and trainingusers are calculated based on context similarity and userinterest similarity

Step 3 Based on the similar values a set of similar users areidentified

Step 4 The basic service ranking results are obtained bytaking advantages of the past service usage experiences ofsimilar users

Step 5 Relational degree among cloud services is evaluatedThrough utilizing relational degree to improving the accuracyof ranking results we succeed in obtaining the ultimateservice ranking results

Step 6 SSRM only selects trusted mobile cloud services ascandidates based on trust filter module

Step 7 The ranking prediction results are provided to theactive user

This paper targets the research task of accurately makingQoS ranking prediction with consideration of user contextinformation and the relation among cloud services accordingto the requirements of different application scenarios Thispaper makes the following contributions

(1) This paper focuses on the trustworthy service selec-tion and recommendation inmobile cloud computingenvironments We propose a novel service selectionand recommendation model (SSRM) where similar-ity is calculated based on user context informationand interest

(2) SSRM calculates the relational degree among servicesbased on PropFlow algorithm and utilizes relationaldegree to improve the accuracy of ranking results

(3) SSRM supports a personalized and trusted selectionof cloud services through taking into account mobileuserrsquos trust expectation

(4) Simulation experiments are conducted on ns3 simu-lator to study the prediction performance of SSRMcompared with other two traditional approachesitem-based CF approach (IBCF) and user-based CFapproach (UBCF)The experimental results show theeffectiveness of SSRM

This paper is organized as follows Section 2 describesrelated work In Section 3 the proposed SSRM is discussedSection 4 describes the test scenario and simulation resultsFinally we conclude with a summary of our results anddirections for new research in Section 5

2 Related Work

Mobile cloud computing (MCC) has attracted extensiveattention in recent years since it provides real-time accessto real-time information through the applications on mobiledevices The rapid advancements of mobile cloud technologyhave been the prime reason for significant expansions in thismarket where there exist abundant functionally similar ser-vices provided by cloud vendors including Amazon GoogleInc Apple Inc and Microsoft Corporation Markets andMarkets [11] forecast the global mobile cloud market willgrow to over $4690 Billion by 2019 The exponential growthof mobile cloud market makes it hard for users to select

4 Wireless Communications and Mobile Computing

the most suitable mobile cloud service Therefore how toselect trustworthy and high quality mobile cloud servicesbecomes one of the most urgent problems A number ofworks have been carried out on cloud service selection issueincluding rating-oriented collaborative filtering methods [45] ranking-oriented collaborative filtering methods [3] andsome other methods

Rating-oriented or ranking-oriented collaborative filter-ing methods can produce QoS prediction or service rank-ing using collaborative filtering technology Rating-orientedcollaborative filtering methods rank the cloud services basedon the predicted QoS values QoS presents the nonfunctionalproperties of cloud services including security availabilitythroughput and response-time properties Based on theservice QoS measures various approaches are proposed forservice selection Pan et al [4] propose a trust-enhancedcloud service selection model based on QoS analysis In themodel trust is utilized to find similar neighbors and predictthemissingQoS values Ding et al [5] propose a personalizedcloud service selection method to make experience usabilityand value distribution to measure the service similarityRating-oriented collaborative filtering approaches first pre-dict the missing QoS values before making QoS rankingThe target of rating-oriented approaches is to predict QoSvalues as accurate as possible However accurate QoS valueprediction may not lead to accurate QoS ranking prediction[3]

Compared with rating-oriented methods ranking-oriented methods predict the QoS rankings directlyDifferent from the previous ranking-oriented methodsZheng et al [3] propose a comprehensive study of how toprovide accurate QoS ranking for cloud services Althoughranking-oriented methods can be used to make optimalcloud service selection from a set of functionally equivalentservice candidates these methods ignore the changes ofQoS QoS in MCCmight vary largely even for the same typeof mobile cloud services QoS in MCC are affected by theuser context information The rating-oriented collaborativefilteringmethods and ranking-oriented collaborative filteringmethods both can help user to predict missing QoS valuebut they did not consider the dynamic QoS properties inMCC

Other types of service selection approaches are alsowidely examined The most employed approaches includemulticriteria decision analysis-based service selection [7 12ndash14] reputation-aware service selection [15] adaptive learn-ing mechanism-based service selection [8 16] economictheoretical model-based service selection [17 18] servicelevel agreement-based service ranking [6] visualizationframework for service selection [19] and trust evaluationmiddleware for cloud service selection [20] Though theseapproaches can efficiently measure service quality the imple-mentation of some approaches is time-consuming and costly

Ma et al [7] propose a time-aware service selectionapproach by using interval neutrosophic set In the paper thestrategy for selecting trustworthy services from an abundantfield of candidates involves formulating the problem of time-aware service selection with tradeoffs between performance-costs and potential risk as a multicriterion decision-making

(MCDM) problem that creates a ranked services list usinginterval neutrosophic set (INS) theory The experimentalresults demonstrate that the proposed approach can workeffectively in both the risk-sensitive service selection modeand the performance-cost-sensitive service selection modebut the approach ignores the changes of user context infor-mation in MCC environments

Whaiduzzaman et al [12] identify and synthesize severalmulticriteria decision analysis (MCDA) techniques and pro-vide a comprehensive analysis of this technology for generalreaders In addition this paper presents a taxonomy derivedfrom a survey of the current literature The results show thatMCDA techniques are indeed effective and can be used forcloud service selection but that different techniques do notselect the same service

This paper [14] presents a cloud service selectionmethod-ology that utilizesQuality-of-Service history of cloud servicesover different time periods and performs parallel multicri-teria decision analysis to rank all cloud services in eachtime period in accordance with user preferences beforeaggregating the results to determine the overall rank ofall the available options for cloud service selection Thismethodology assists the cloud service user to select the bestpossible available service according to the requirements Themain disadvantages are that the framework proposed in thispaper deals with service selection in the preinteraction periodonly Work on postinteraction service migration decisions isneeded and several other important factors such as the costof migration in terms of service disruption and data transferalso need to be included in the decision-making process

Fan et al [15] propose a multidimensional trust-awarecloud service selection mechanism based on evidential rea-soning approach which aggregates multidimensional trustfeedback ratings to form the reputation values of the cloudservice providersThough the experimental results show thatthis approach is effective in cloud systems it can be only usedto recommend an optimal cloud service which satisfies thekey QoS requirements for users

Wang et al [16] propose a dynamic cloud service selectionmethod by using an adaptive learning mechanism whichinvolves incentive forgetting and degenerate functions thatcan realize the self-adaptive regulation for optimizing nextservice selection according to the status of current serviceselection To assist users to efficiently select their preferredcloud services a cloud service selection model adopting thecloud service brokers is given in the paper However inthe paper the brokerage scheme on cloud service selectiontypically assumes that brokers are completely trusted and donot provide any guarantee over the correctness of the servicerecommendations It is then possible for a compromisedor dishonest broker to easily take advantage of the limitedcapabilities of the clients and provide incorrect or incompleteresponses

Do et al [17] propose a dynamic service selectionmethodwhich provides a price game in heterogeneous cloud marketThe paper studies price competition in a heterogeneous cloudmarket where users can identify benefit and cost of cloudservice application and choose the best one through ana-lyzing market-relevant factors But this paper only considers

Wireless Communications and Mobile Computing 5

one service In real-world applications there are many cloudservices in the practical cloudmarket Additionally this paperhas not considered service level agreement issues that are alsoimportant for cloud users

The visualization framework for service selection [19]takes into cognizance the set of cloud services that matchesa userrsquos request and based on QoS attributes users caninteract with the results via bubble graph visualization tocompare and contrast the search results to ascertain the bestalternative Although the result from the experiments showsthat visualization framework simplifies decision-making theuse of bubble graph introduces additional complexity for theuser when making a suitable service selection

The authors [20] discuss themethod of enhancing servicetrust evaluation and propose a trustworthy selection frame-work for cloud service selection named TRUSS Aimingat developing an effective trust evaluation middleware forTRUSS this paper proposes an integrated trust evalua-tion method via combining objective trust assessment andsubjective trust assessment Simulation-based experimentsvalidated the performance of the proposed method but themethod assumes that the majority of service users are honestand a dishonest user gives more unfair ratings than fairratings This assumption is unrealistic

Cloud computing and mobile cloud computing are quitedifferentThemobile cloudpays attention to services availablethrough mobile network operators (MNOs like Verizon andATampT) While a great number of researchers have focusedon the service selection and recommendation in cloudcomputing little attention has been devoted to trustworthyservice selection in mobile cloud computing Different fromthese existing approaches our work focuses on how to selecttrustworthy and high quality cloud services formobile user inMCC which is an urgently required research problemThuswe propose a context-aware cloud service selection methodfor mobile cloud computing environments which considersuser context information and interest in order to calculateuser similarities and utilize relational degree among cloudservices to improve the accuracy of service ranking results

3 SSRM

In this section we present a detailed discussion of theproposed novel service selection and recommendationmodel(SSRM) Firstly user similarities are calculated and thebasic personalized service ranking results are obtained inSection 31 Secondly relational degree is calculated and usedto improve the accuracy of ranking results in Section 32

31 User Similarity Computation Given119898users and 119899mobilecloud services the user-service matrix (USM) for predictingthe missing service usage experiences is denoted as

[[[[1199031199061 mcs1 sdot sdot sdot 1199031199061 mcs119899 d

119903119906119898 mcs1 sdot sdot sdot 119903119906119898 mcs119899

]]]] (1)

where 119903119906119898 mcs119899 expresses historical service usage experiences(QoS value or customer rating) of mobile cloud service mcs119899made by user 119906119898 ldquo119903119898119899 = nullrdquo states that 119906119898 did notinvoke mcs119899 yet In order to get the ranking results we firstlycalculate user similarityThe user similarity is measured withthe following equation

sim119880 (119906119909 119906119910) = 120582 lowast sim119862 (119906119909 119906119910)+ (1 minus 120582) sim119868 (119906119909 119906119910) (2)

where 119906119909 and 119906119910 denote two users sim119862(119906119909 119906119910) andsim119868(119906119909 119906119910) respectively express context similarity and userinterest similarity between 119906119909 and 119906119910 120582 is defined to deter-mine how much similarity measure relies on context andinterest 120582 is in the interval [0 1]311 Context Similarity Computation Cloud service selec-tion of a user in a context (eg a laptop with enough battery)may bemuch different from that of another user in a differentcontext (eg a smart phone without enough battery) Hencecloud service selection can be significantly affected by usercontext To evaluate the context similarities between targetuser and training user we provide the definition of contextThe definition from [21] is referenced by us Let 119903119906119909 mcs119895 beusage experience of cloud service mcs119895 made by user 119906119909Definition 1 (context) Context 119862119906119909 mcs119895 is any informationthat can be used to characterize the situation where usageexperience 119903119906119909 mcs119895 of cloud service mcs119895 is made by user 119906119909

Context information has different properties whichinclude location resource ability and status Let V be the typeof context property of 119862119906119909 mcs119895

The context 119862119906119909 mcs119895 is denoted as

119862119906119909 mcs119895 = (119862nor119906119909mcs119895 (1) 119862nor

119906119909mcs119895 (2) 119862nor119906119909mcs119895 (V)) (3)

where 119862nor119906119909mcs119895(V) is the normalized property value

Pearson Correlation Coefficient (PCC) has been success-fully employed to obtain the numerical distance betweendifferent users for similarity calculation [4 5] Let 119906119909 and 119906119910be two users then PCC is applied to calculate the contextsimilarity between 119906119909 and 119906119910 over service mcs119895 by

sim119862mcs119895(119906119909 119906119910) = sumV

119904=1 (119862nor119906119909 mcs119895 (119904) minus 119862nor

119906119909mcs119895 (119904)) (119862nor119906119910mcs119895 (119904) minus 119862nor

119906119910mcs119895 (119904))radicsumV119904=1 (119862nor

119906119909mcs119895 (119904) minus 119862nor119906119909mcs119895 (119904))2radicsumV

119904=1 (119862nor119906119910mcs119895 (119904) minus 119862nor

119906119910 mcs119895 (119904))2 (4)

6 Wireless Communications and Mobile Computing

where sim119862mcs119895(119906119909 119906119910) expresses the context similarity

between 119906119909 and 119906119910 and it is in the interval of [minus1 1] and119862119906119909 mcs119895(119904) and 119862119906119910 mcs119895(119904) stand for the average values ofcontext

312 Interest Similarity Computation In mobile cloud com-puting environments each user lives in a large network offriends which is called social network It has been reportedthat user interestsrsquo similarity can be leveraged to supportservices and products recommendation [22] A user prefersto choose the items recommended by other users withsimilar interest in a social network Exploring those usersof a high interest similarity with the existing clients couldefficiently enlarge client groups for cloud service providers[22]

Here we use cosine similarity to compute interest similar-ity similar to [22] Interest similarity between users 119906119909 and 119906119910is then defined as the cosine distance between their respectivecloud service invocation sets

sim119868 (119906119909 119906119910) = 100381710038171003817100381710038171003817119868119906119909 cap 1198681199061199101003817100381710038171003817100381710038171003817100381710038171003817100381711986811990611990910038171003817100381710038171003817 sdot 100381710038171003817100381710038171003817119868119906119910100381710038171003817100381710038171003817 (5)

where 119868119906119909 = radic119897119906119909 (119897119906119909 is the number ofmobile cloud serviceswhich have been invoked by 119906119909) and 119868119906119909 cap119868119906119910 is the numberof mobile cloud services which have been coinvoked by 119906119909and 119906119910 If 119897119906119909 = 0 or 119897119906119910 = 0 sim119868(119906119909 119906119910) is undefined

Cloud service invocation is an important factor to deter-mine the interest of users [5] For example user 119906119909 hasinvoked mobile cloud services mcs1 mcs3 mcs6 and mcs7119906119910 has invoked mcs6 and mcs7 and 119906119911 has invoked mcs1 andmcs2 Though 119906119909 119906119910 and 119906119911 do not know each other in realworld and 119906119909 and 119906119910 have a high interest similarity than 119906119909and 119906119911 as they both invoked mcs6 and mcs7 [4]

313 Getting Initial Ranking Results We first predict themissing value 119903119906119909 mcs119895 of mobile cloud service mcs119895 madeby user 119906119898 before making initial ranking results For everymobile cloud service only the Top-119896 similar training useris selected to make missing value prediction Based on thesim119880(119906119909 119906119910) values a set of the Top-k similar training users119873119862mcs119895

(119906119909) over service mcs119895 is identified for target user 119906119909by

119873119862mcs119895(119906119909)

= 119906119910 | 119906119910 isin TS119906119909 sim119880 (119906119909 119906119910) gt 0 119906119909 = 119906119910 (6)

where TS119906119909 is a set of the Top-k similar training users to targetuser 119906119909 and sim119880(119906119909 119906119910) gt 0 excludes the dissimilar userwith negative similar values The value of sim119880(119906119909 119906119910) in (6)is calculated by (2) The missing value 119903119906119909mcs119895 is predicted asfollows

119903119906119909mcs119895= 119903119906119909+ sum119906119910isin119873119862mcs119895

(119906119909)(119903119906119910 mcs119895 minus 119903119906119910) sim119880 (119906119909 119906119910)sum119906119910isin119873119862mcs119895(119906119909)

sim119880 (119906119909 119906119910) (7)

where 119903119906119909 and 119903119906119910 are the average value of service usageexperiences (QoS value or customer rating) of mobile cloudservice made by users 119906119909 and 119906119910 respectively As differentQoS properties have different dimensions and range of valueswe first ensure predicted missing values in the range of [0 1]In [5] QoS properties are classified into two categories ldquocostrdquoand ldquobenefitrdquo For ldquocostrdquo property (response-time) the lowerits value is the greater possibility that a user would chooseit becomes In SSRM all QoS properties are consideredas ldquobenefitrdquo attribute Normalized missing value 119876119906119909 mcs119895 iscomputed as follows

119876119906119909 mcs119895

=

119903119906119909mcs119895 minus min (119903119906119909)max (119903119906119909) minus min (119903119906119909) 119903119906119909mcs119895 isin ldquobenefitrdquo

max (119903119906119909) minus 119903119906119909mcs119895

max (119903119906119909) minus min (119903119906119909) 119903119906119909mcs119895 isin ldquocostrdquo(8)

where min(119903119906119909) and max(119903119906119909) denote the minimum andmaximumQoSproperty value for user119906119909 and they are subjectto the following constraints

min (119903119906119909) = min (119903119906119909mcs119895 | 119895 = 1 119899)max (119903119906119909) = max (119903119906119909 mcs119895 | 119895 = 1 119899) (9)

119876119906119909 mcs119895 is in the range of [0 1] The larger its value isthe more possibility that user 119906119909 would be satisfied with theservice becomes Finally the cloud services are ranked for 119906119909in the order of decreasing 119876119906119909mcs119895 values for cloud serviceselection similar to [4]The basic ranking algorithm is shownin Algorithm 1 The basic ranking algorithm is similar to [3]Compared with [3ndash5] in our ranking algorithm the usersimilarity is measured based on user context informationand user interest In addition the basic ranking algorithm isenhanced in Section 32 In the enhanced ranking algorithmwe calculate the relational degree among cloud services andutilize it to improve the accuracy of ranking results

The basic algorithm ranks the employed mobile cloudservice in 119864 based on the observed historical service usageexperiences The ranking results are stored in 119877119890(119905) which isin the interval [1 |119864|] A smaller value shows a higher rank119905 expresses a mobile cloud service For every cloud servicemcs119895 normalized value 119876119906119909 mcs119895 is calculated Services areranked from high to low by picking up the service 119905 that hasthe maximum 119876119906119909mcs119895 value

Wireless Communications and Mobile Computing 7

Inputan employed service set 119864a full service set 119868an employed similar training user set 119873119862mcs119895

(119906119909)observed service usage experience values 119903119906119910 mcs119895 (119906119910 isin 119873119862mcs119895

(119906119909))Outa service ranking 119865 = 119864while 119865 = Oslash do119905 = argmaxmcs119895isin119865119876119906119909 mcs119895 119877119890(119905) = |119864| minus |119865| + 1119865 = 119865 minus 119905endforeach mcs119895 isin 119868 do

119876119906119909 mcs119895 =

119903119906119909 mcs119895 minus min(119903119906119909 )max(119903119906119909 ) minus min(119903119906119909 ) 119903119906119909 mcs119895 isin ldquobenefitrdquo

max(119903119906119909 ) minus 119903119906119909 mcs119895

max(119903119906119909 ) minus min(119903119906119909 ) 119903119906119909 mcs119895 isin ldquocostrdquo

end119899 = |119868|while 119868 = Oslash do119905 = argmaxmcs119895isin119868119876119906119909 mcs119895(119905) = 119899 minus |119868| + 1119868 = 119868 minus 119905end

end

Algorithm 1 Basic ranking algorithm

32 Improving Accuracy of Ranking Results Through calcu-lating the similar relation among users the missing value119903119906119909mcs119895 is successfully predicted in Section 31 However therelation among cloud services is no doubt an importantevaluation factor of missing value prediction The relatedservices will be probably selected by the same user based onthe intuition [9]Therefore we calculate the relational degreeand utilize it to improve the accuracy of missing value 119903119906119909mcs119895computation in the section The relational degree betweenservice mcs119895 and service mcs119894 is denoted as 119889119895119894 In the paperPropFlow [23] is used to calculate 119889119895119894 PropFlow algorithmcomputes the information flow between services where alarger value indicates tighter relation In order to calculate119889119895119894 we firstly describe the formal definition of service relationgraph

Definition 2 (service relation graph) Given a set of servicesMCS and a totally ordered domain of weights 119882 a servicerelation graph is a weighted undirected graph 119866 = (MCS 119864)The edge (mcs119894mcs119895 119908119894119895) in the set 119864 isin 119880 times 119880 times 119882 encodesthe link weight119908119894119895 (119908119894119895 ge 0) between service mcs119894 and servicemcs119895 An edge depicts the direct link relationship betweenmcs119894 and mcs119895 As 119866 is undirected 119908119894119895 = 119908119895119894

Equation (10) shows how to calculate 119889119895119894119889119895119894 = Input119895 lowast 119908119895119894sum119896isin119873(119895) 119908119895119896 (10)

Ｇ＝Ｍℎ

Ｇ＝ＭiＧ＝Ｍl

Ｇ＝Ｍe

Ｇ＝Ｍf

Ｇ＝Ｍj Ｇ＝Ｍk

wkl = 1 wli = 5

wle = 1

wie = 1wke = 1

wkℎ = 2

wjk = 3

wjf = 1

Figure 3 An example of service relation graph

where initial input is regarded as 1 and 119873(119895) is the set ofneighbors of mcs119895

Figure 3 shows an example of relation degree computa-tion where there are six kinds of mobile cloud services mcs119894and mcs119895 are indirectly linked We assume mcs119895 is startingnode

Four paths can reach mcs119894 from mcs119895 but only twopaths are considered for computation based on PropFlow

8 Wireless Communications and Mobile Computing

algorithm They are mcs119895 rarr mcs119896 rarr mcs119897 rarr mcs119894 andmcs119895 rarr mcs119896 rarr mcs119890 rarr mcs119894

First we compute 119889119895119896 The sum of link weights of mcs119895and its neighbor is 4 119889119895119896 is computed as

119889119895119896 = 1 lowast 3(1 + 3) = 1 lowast 34 = 34 (11)

119889119896119897 is computed as

119889119896119897 = 119889119895119896 lowast 1(1 + 1 + 2) = 34 lowast 14 = 316 (12)

With the same method 119889119895119894 is computed as

119889119895119894 = 119889119897119894 + 119889119890119894 = 119889119896119897 lowast 55 + 119889119896119890 lowast 11 = 38 (13)

More details of calculating 119889119895119894 are shown in [24]Next we assume service mcs119894 is invoked recently by user119906119909 119889119895119894 is incorporated into (7) and then the missing value119903119906119909mcs119895 is computed with (14) in basic ranking algorithm

Finally the ultimate ranking results are got

119903119906119909 mcs119895 = (119903119906119909+ sum119906119910isin119873119862mcs119895

(119906119909)(119903119906119910 mcs119895 minus 119903119906119910) sim119880 (119906119909 119906119910)sum119906119910isin119873119862mcs119895(119906119909)

sim119880 (119906119909 119906119910) )sdot (1 + 119889119895119894)

(14)

In SSRM we only select trusted mobile cloud serviceas candidates Therefore a set of trusted candidates areidentified for 119906119909 by

MCS119909 = mcs119895 | trustmcs119895 gt 120583119909 119895 = 1 119873 (15)

where trustmcs119895 denotes the trustworthiness of mobile cloudservice mcs119895 and 120583119909 expresses the trust threshold decidedby user 119906119909 Many existing methods [4ndash6] can be applied tocompute trustmcs119895 Here how to compute trustmcs119895 is not tobe discussed We omit the details for brevity

4 Experimental Study

In this section in order to evaluate the effectiveness ofSSRM a series of test scenarios are developed To studythe prediction performance we compare SSRM with othertwo traditional approaches item-based CF approach (IBCF)and user-based CF approach (UBCF) SSRM IBCF andUBCF are rating-oriented methods which rank the cloudservices based on the predicted QoS values Since there isno suitable real data supporting mobile cloud computingsimulation we use ns3 simulator to generate the experimentaldataset Firstly we generated cloud service invocation codes

by Axis2 [25] a Java-based open source package for cloudservices Then we simulate mobile cloud service on ns3[26] based on the invocation codes On ns3 the creationof servers mobile users and service model was easier thanwith other network simulators In the simulating process ofdataset the real-world web service QoS dataset from WS-DREAM team [3] is referenced by usThe detailed real-worldQoS values are publicly released online [27] which makesour experimental evaluations reproducible We simulate 100mobile users 25 servers and 300 services The value of linkweight between two services is randomly generated which isin the range of [1 5] In order to conduct our experimentsrealistically the changes of user context information aresimulated Three types of context are considered includingbandwidth memory and battery capacity The bandwidthvalue of a mobile user is in the range of 4Mndash8MThe valuesof memory and battery capacity will decline over timeThesevalues are in the range of 100000000ndash2100000000MB and05ndash2KVA

We normalize the context information by

119862nor119906119909mcs119895 (V) = 119886 tan (119862119906119909 mcs119895 (V)) lowast 2120587 (16)

where 119862nor119906119909mcs119895(V) denotes the normalized context value of

user 119906119909 over cloud service mcs119895 Inverse tangent function isapplied to normalize the raw data 119862119906119909 mcs119895(V) Then thesenormalized property values are used to calculate contextsimilarity with (4) Once the context simulation is finishedthe target user 119906119909 will obtain context similarity results foreach cloud service

Similar to [3] QoS properties include throughput andresponse-time that are in the range of 0ndash1000 kbps and 0ndash20second respectively Most of the throughput and response-time values fall between 5ndash40 kbps and 01ndash08 seconds Inorder to simply the experimental process 200 service QoSrecords are randomly selected and two 100 times 200 user-servicematrices are constructed The two matrices include through-put and response-time respectivelyUser-servicematrices areseparated into two parts training set (80 historical usageexperiences in the matrix) and test set (the remaining 20usage experiences) Each entry in the matrix is the QoSvalue (eg response-time or throughput) of a mobile cloudservice observed by a user The experiments employ the QoSvalues of response-time and throughput to rank the servicesindependently Table 1 shows descriptions of the obtainedpractical cloud service QoS values and context information

We use Mean Absolute Error (MAE) and Root MeanSquare Error (RMSE) as the metric to evaluate predictionperformance of the proposed approach in comparison withother approaches MAE and RMSE are defined as

MAE = sum119906119909 mcs119895

100381610038161003816100381610038161003816119903act119906119909mcs119895 minus 119903pre119906119909mcs119895100381610038161003816100381610038161003816119897pre

RMSE = radic sum119906119909 mcs119895 (119903act119906119909mcs119895 minus 119903pre119906119909 mcs119895)2119897pre (17)

Wireless Communications and Mobile Computing 9

Response-time

02

025

03

035

04

045

05

055

06

065

07

MA

E

01 02 03 04 05 06 07 08 09 100Weighting factor

(a) MAE values

Response-time

03

035

04

045

05

055

06

065

07

075

08

RMSE

01 02 03 04 05 06 07 08 09 100Weighting factor

(b) RMSE values

Figure 4 Impact of weighting factor 120582

Table 1 Mobile cloud service QoS dataset and context informationdescriptions

Statistics ValueNumber of mobile cloud serviceinvocations 180000

Number of service users 100Number of mobile cloud services 200Minimum response-time value 0004 sMaximum response-time value 20 sMean of response-time 110 sStandard deviation ofresponse-time 226 s

Minimum throughput value 01 kbpsMaximum throughput value 1000 kbpsMean of throughput 2546 kbpsStandard deviation of throughput 4803 kbpsBandwidth 4Mndash8MMemory 100000000MBndash2100000000MBBattery capacity 05 KVAndash2KVA

where 119903act119906119909mcs119895 and 119903pre119906119909mcs119895 denote the actual QoS valueand predicted value respectively 119897pre is the number ofpredicted QoS values Smaller values of MAE and RMSEindicate better results

41 Impact of 120582 120582 is weighting factor in (2) 120582 determineshow much similarity measure relies on context and interestIn the experiment we vary the weighting factor 120582 from 0 to 1in increment of 01 The Top-k is set to 5 Matrix density is setto 10 Matrix density means the percentage of selected QoSentries that are used to predict the missing QoS value Theexperimental results are shown in Figure 4 As the value of 120582increases MAE firstly decreases and then quickly increasesWhen 120582 = 04 our results have the best value As shown inFigure 4(b) RMSE presents similar trend

42 Performance Comparison of SSRM IBCF and UBCFTo compare the performance of SSRM IBCF and UBCFsimilarity computation methods we implement IBCF andUBCF methods

Item-Based CF Approach (IBCF) We apply PCC to calculatesimilarities between users and predicted QoS values based onsimilar users The user similarity is calculated by

sim (119906119909 119906119910) = summcs119895isinmcs119906119909119906119910(119903119906119909 mcs119895 minus 119903119906119909) (119903119906119910 mcs119895 minus 119903119906119910)

radicsummcs119895isinmcs119906119909119906119910(119903119906119909mcs119895 minus 119903119906119909)2radicsummcs119895isinmcs119906119909119906119910

(119903119906119910mcs119895 minus 119903119906119910)2 (18)

10 Wireless Communications and Mobile Computing

where mcs119906119909119906119910 is the set of mobile cloud services that havebeen coinvoked by 119906119909 and 119906119910 119903119906119909and 119903119906119910 are average QoSvalues of cloud service invoked by 119906119909 and 119906119910 respectively

User-Based CF Approach (UBCF) We apply PCC to calculatesimilarities between services and predicted QoS values basedon similar services The service similarity is calculated by

sim (mcs119895mcs119894) = sum119906119909isin119906mcs119895mcs119894(119903119906119909mcs119895 minus 119903mcs119895) (119903119906119909 mcs119894 minus 119903mcs119894)

radicsum119906119909isin119906mcs119895mcs119894(119903119906119909mcs119895 minus 119903mcs119895)2radicsum119906119909isin119906mcs119895mcs119894

(119903119906119909mcs119894 minus 119903mcs119894)2 (19)

where 119906mcs119895mcs119894 is the set of users who have invoked the cloudservices mcs119895 andmcs119894 119903mcs119895 and 119903mcs119894 are average QoS valuesof cloud services mcs119895 and mcs119894 made by 119906119909

In the experiments trustworthiness of cloud service hasno influence on similarity computation and service decisionin the experiments In order to simplify the experimentalprocess we consider all cloud services are trustworthyWeighting factor 120582 is set to 04 We firstly study the effectof neighbor size 119896 Matrix density is set to 10 We change119896 from 5 to 30 in increment of 5 Figure 5 shows theexperimental results for response-time and throughput

As shown in Figure 5 the prediction performance ofSSRM outperforms other two approaches The performanceof IBCF is similar to UBCF As the values of 119896 increase betteraccuracy can be achieved This indicates that more similarneighbor records can provide more information for missingvalue prediction However when 119896 is larger than 25 MAEand RSME fail to drop with the value of 119896 increasing Themain reason is the limited number of similar neighbors Theobservations also show that RMSE has the similar trend butwith larger fluctuations

Next we investigate the impact of matrix density Thematrix density varies from 10 to 50 in increment of10 Similar user size k is set to 5 in this experimentFigure 6 shows the experimental results for response-timeand throughput

As presented in Figure 6 the prediction performance isalso enhanced with the value of matrix density increasingThe main reason is that denser user-service matrix providesmore information for missing value prediction NARM hasthe better performance than LBCF and UBCF under allexperimental setting consistently since NARM considerscontext similarity and the relational degree among cloudservices The experimental results of LBCF and UBCF aresimilar in this experiment since these two similarity compu-tation methods are similar to each other and are both rating-oriented

43 Performance Comparison with Other Popular ApproachesTo study the prediction performance of SSRM we com-pare SSRM with three existing QoS properties predictionapproaches CloudRank2 [3] TECSS [4] and JV-PCC [5]

The size of Top-119896 similar service is an important factorin CF approach which determines how many neighborsrsquohistorical records are employed to generate predictions [5]Therefore the impact of matrix density is not considered in

the experiments We set the density to 10 We vary 119896 from 5to 30 in increment of 5 In order to simplify the experimentalprocess we consider all cloud services are trustworthyFigure 7 shows the experimental results for response-timeand throughput Under the same simulation condition SSRMsignificantly outperforms CloudRank2 TECSS and JV-PCCThe observations also suggest that better accuracy can beachieved by our model when more historical records areavailable in the service selection studyThemain reason is thatcomputing context similarity can improve the performance ofprediction evaluation in MCC environments

Figures 7(a) and 7(b) depict the MAE fractions of SSRMCloudRank2 TECSS and JV-PCC for response-time andthroughput while Figures 7(c) and 7(d) depict the RMSEfractions It can be observed that SSRM achieves smallerMAE and RMSE consistently than other approaches for bothresponse-time and throughput

5 Conclusions and Future

In the paper we propose a novel service selection and recom-mendation model (SSRM) for mobile cloud computing envi-ronmentsWhen calculating user similarity context informa-tion and interest are considered Through utilizing relationaldegree to improve the accuracy of ranking result our serviceselection and recommendation method succeed in obtainingthe ultimate service ranking results In addition SSRM onlyselects trusted mobile cloud service as candidates Thereforea set of trusted candidates are identified for target user Theexperimental results show that our approach significantlyimproves the prediction performance as comparedwith othertwo traditional approaches item-based CF approach (IBCF)and user-based CF approach (UBCF)

There are some disadvantages in our SSRM Firstly weexploit node distance to compute context similarity Thereis an underlying assumption in this exploitation each ofthe two adjacent nodes has equal semantic distance orgranularity of nodes in each level is identicalThis underlyingassumption is not true in some scenarios and it deterioratesthe performance of similarity evaluation Secondly the trustthreshold is used to select trusted mobile cloud service ascandidates However it cannot detect and exclude maliciousQoS values provided by users Thirdly in our experimentsthere are only three types of context information that areconsidered

Therefore in the future we will study the methods toimprove context similarity measurement We will study how

Wireless Communications and Mobile Computing 11

Response-time

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E10 25 3015 205

k

(b)Response-time

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

(c)

Throughput

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

(d)

Figure 5 Impact of neighbor size 119896 comparison with SSRM IBCF and UBCF

12 Wireless Communications and Mobile Computing

Response-time

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E

20 50 6030 4010Matrix density ()

(a)

Throughput

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E20 30 40 50 6010

Matrix density ()

(b)

Response-time

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 50 6030 4010Matrix density ()

(c)

Throughput

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 30 40 50 6010Matrix density ()

(d)

Figure 6 Impact of matrix density comparison with SSRM IBCF and UBCF

Wireless Communications and Mobile Computing 13

Response-time

JV-PCCTECSSSSRM

CloudRank2

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

02

025

03

035

04

045

05

055

06

065

07

MA

E10 15 20 25 305

k

JV-PCCTECSSSSRM

CloudRank2

(b)

Response-time

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

JV-PCCTECSSSSRM

CloudRank2

(c)

Throughput

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

JV-PCCTECSSSSRM

CloudRank2

(d)

Figure 7 Impact of neighbor size 119896 comparison with other popular approaches

14 Wireless Communications and Mobile Computing

to select themost trustworthy cloud service of certain type forthe active users based on multidimensional trust evidenceWe also will conduct more experimental investigations thatdeal with the impact of different context changes on serviceselection Moreover we will improve the ranking accuracyof our approaches by exploiting additional techniques (com-bining rating-oriented approaches and ranking-orientedapproaches matrix factorization intelligent optimizationalgorithms etc)

Conflicts of Interest

The author declares that they have no conflicts of interest

Acknowledgments

The work in this paper has been supported by NationalNatural Science Foundation of China (Program no 71501156)and China Postdoctoral Science Foundation (Program no2014M560796) and Shanxi Provincial EducationDepartment(Program no 15JK1679)

References

[1] White PaperMobile CloudComputing Solution Brief AEPONA2010

[2] httpsenwikipediaorgwikiCloudlet[3] Z Zheng X Wu Y Zhang M R Lyu and J Wang ldquoQoS

ranking prediction for cloud servicesrdquo IEEE Transactions onParallel and Distributed Systems vol 24 no 6 pp 1213ndash12222013

[4] Y Pan SDingW Fan J Li and S Yang ldquoTrust-enhanced cloudservice selection model based on QoS analysisrdquo PLoS ONE vol10 no 11 Article ID e0143448 2015

[5] S Ding C-Y Xia K-L Zhou S-L Yang and J S Shang ldquoDeci-sion support for personalized cloud service selection throughmulti-attribute trustworthiness evaluationrdquo PLoS ONE vol 9no 6 Article ID e107821 2014

[6] S K Garg S Versteeg and R Buyya ldquoA framework for rankingof cloud computing servicesrdquo Future Generation ComputerSystems vol 29 no 4 pp 1012ndash1023 2013

[7] H Ma Z Hu K Li and H Zhang ldquoToward trustworthy cloudservice selection a time-aware approach using interval neutro-sophic setrdquo Journal of Parallel and Distributed Computing vol96 pp 75ndash94 2016

[8] A Ahmed and A S Sabyasachi ldquoMobile cloud service selectionusing back propagation neural networkrdquo International Journalof Computer Applications vol 129 no 14 pp 1ndash5 2015

[9] L Guo J Ma Z-M Chen and H-R Jiang ldquoIncorporatingitem relations for social recommendationrdquo Chinese Journal ofComputers vol 37 no 1 pp 219ndash228 2014

[10] Z Yang B Wu K Zheng X Wang and L Lei ldquoA survey ofcollaborative filtering-based recommender systems for mobileinternet applicationsrdquo IEEE Access vol 4 pp 3273ndash3287 2016

[11] Markets and Markets httpwwwmarketsandmarketscomPressReleasesmobile-cloudasp

[12] M Whaiduzzaman A Gani N B Anuar M Shiraz MN Haque and I T Haque ldquoCloud service selection usingmulticriteria decision analysisrdquoTheScientificWorld Journal vol2014 Article ID 459375 10 pages 2014

[13] K Tserpes F Aisopos D Kyriazis and T Varvarigou ldquoArecommender mechanism for service selection in service-oriented environmentsrdquo Future Generation Computer Systemsvol 28 no 8 pp 1285ndash1294 2012

[14] Z U Rehman O K Hussain and F K Hussain ldquoParallelcloud service selection and ranking based on QoS historyrdquoInternational Journal of Parallel Programming vol 42 no 5 pp820ndash852 2014

[15] W-J Fan S-L YangH Perros and J Pei ldquoAmulti-dimensionaltrust-aware cloud service selection mechanism based on evi-dential reasoning approachrdquo International Journal of Automa-tion and Computing vol 12 no 2 pp 208ndash219 2015

[16] X G Wang J Cao and Y Xiang ldquoDynamic cloud serviceselection using an adaptive learning mechanism in multi-cloudcomputingrdquo The Journal of Systems and Software vol 100 pp195ndash210 2015

[17] C T Do N H Tran E-N Huh C S Hong D Niyato andZ Han ldquoDynamics of service selection and provider pricinggame in heterogeneous cloud marketrdquo Journal of Network andComputer Applications vol 69 pp 152ndash165 2015

[18] R Pal and P Hui ldquoEconomic models for cloud service marketspricing and capacity planningrdquo Theoretical Computer Sciencevol 496 pp 113ndash124 2013

[19] A Ezenwoke O Daramola and M Adigun ldquoTowards avisualization framework for service selection in cloud e-marketplacesrdquo in Proceedings of the 2017 IEEE InternationalConference on Web Services (ICWS rsquo17) pp 828ndash835 HonoluluHI USA June 2017

[20] M Tang X Dai J Liu and J Chen ldquoTowards a trust evaluationmiddleware for cloud service selectionrdquo Future GenerationComputer Systems vol 74 pp 302ndash312 2017

[21] A K Dey ldquoUnderstanding and using contextrdquo Personal andUbiquitous Computing vol 5 no 1 pp 4ndash7 2001

[22] X Han L Wang S Park A Cuevas and N Crespi ldquoAlikepeople alike interests A large-scale study on interest similarityin social networksrdquo in Proceedings of the 2014 IEEEACM Inter-national Conference on Advances in Social Networks Analysisand Mining (ASONAM rsquo14) pp 491ndash496 August 2014

[23] R N Lichtenwalter J T Lussier and N V Chawla ldquoNewperspectives and methods in link predictionrdquo in Proceedings ofthe 16th ACM SIGKDD International Conference on KnowledgeDiscovery and Data Mining (KDD rsquo10) pp 243ndash252 July 2010

[24] L Munasinghe and R Ichise ldquoLink prediction in social net-works using information flow via active linksrdquo IEICE Transac-tion on Information and Systems vol E96-D no 7 pp 1495ndash1502 2013

[25] Axis httpaxisapacheorgaxis2[26] ns3 httpswwwnsnamorgns3[27] Dataset httpwwwzibinzhengcomtpds

Wireless Communications and Mobile Computing 3

User similaritycomputation

Find similar user

Initial rankingresults

Relational gradecomputation

Ultimate rankingresults

User 1

User 2

User m

Mobile cloud netw

orking infrastructure

Trust filter

Context similarity

Interest similarity

Figure 2 Architecture of SSRM

Step 2 User similarities between target user and trainingusers are calculated based on context similarity and userinterest similarity

Step 3 Based on the similar values a set of similar users areidentified

Step 4 The basic service ranking results are obtained bytaking advantages of the past service usage experiences ofsimilar users

Step 5 Relational degree among cloud services is evaluatedThrough utilizing relational degree to improving the accuracyof ranking results we succeed in obtaining the ultimateservice ranking results

Step 6 SSRM only selects trusted mobile cloud services ascandidates based on trust filter module

Step 7 The ranking prediction results are provided to theactive user

This paper targets the research task of accurately makingQoS ranking prediction with consideration of user contextinformation and the relation among cloud services accordingto the requirements of different application scenarios Thispaper makes the following contributions

(1) This paper focuses on the trustworthy service selec-tion and recommendation inmobile cloud computingenvironments We propose a novel service selectionand recommendation model (SSRM) where similar-ity is calculated based on user context informationand interest

(2) SSRM calculates the relational degree among servicesbased on PropFlow algorithm and utilizes relationaldegree to improve the accuracy of ranking results

(3) SSRM supports a personalized and trusted selectionof cloud services through taking into account mobileuserrsquos trust expectation

(4) Simulation experiments are conducted on ns3 simu-lator to study the prediction performance of SSRMcompared with other two traditional approachesitem-based CF approach (IBCF) and user-based CFapproach (UBCF)The experimental results show theeffectiveness of SSRM

This paper is organized as follows Section 2 describesrelated work In Section 3 the proposed SSRM is discussedSection 4 describes the test scenario and simulation resultsFinally we conclude with a summary of our results anddirections for new research in Section 5

2 Related Work

Mobile cloud computing (MCC) has attracted extensiveattention in recent years since it provides real-time accessto real-time information through the applications on mobiledevices The rapid advancements of mobile cloud technologyhave been the prime reason for significant expansions in thismarket where there exist abundant functionally similar ser-vices provided by cloud vendors including Amazon GoogleInc Apple Inc and Microsoft Corporation Markets andMarkets [11] forecast the global mobile cloud market willgrow to over $4690 Billion by 2019 The exponential growthof mobile cloud market makes it hard for users to select

4 Wireless Communications and Mobile Computing

the most suitable mobile cloud service Therefore how toselect trustworthy and high quality mobile cloud servicesbecomes one of the most urgent problems A number ofworks have been carried out on cloud service selection issueincluding rating-oriented collaborative filtering methods [45] ranking-oriented collaborative filtering methods [3] andsome other methods

Rating-oriented or ranking-oriented collaborative filter-ing methods can produce QoS prediction or service rank-ing using collaborative filtering technology Rating-orientedcollaborative filtering methods rank the cloud services basedon the predicted QoS values QoS presents the nonfunctionalproperties of cloud services including security availabilitythroughput and response-time properties Based on theservice QoS measures various approaches are proposed forservice selection Pan et al [4] propose a trust-enhancedcloud service selection model based on QoS analysis In themodel trust is utilized to find similar neighbors and predictthemissingQoS values Ding et al [5] propose a personalizedcloud service selection method to make experience usabilityand value distribution to measure the service similarityRating-oriented collaborative filtering approaches first pre-dict the missing QoS values before making QoS rankingThe target of rating-oriented approaches is to predict QoSvalues as accurate as possible However accurate QoS valueprediction may not lead to accurate QoS ranking prediction[3]

Compared with rating-oriented methods ranking-oriented methods predict the QoS rankings directlyDifferent from the previous ranking-oriented methodsZheng et al [3] propose a comprehensive study of how toprovide accurate QoS ranking for cloud services Althoughranking-oriented methods can be used to make optimalcloud service selection from a set of functionally equivalentservice candidates these methods ignore the changes ofQoS QoS in MCCmight vary largely even for the same typeof mobile cloud services QoS in MCC are affected by theuser context information The rating-oriented collaborativefilteringmethods and ranking-oriented collaborative filteringmethods both can help user to predict missing QoS valuebut they did not consider the dynamic QoS properties inMCC

Other types of service selection approaches are alsowidely examined The most employed approaches includemulticriteria decision analysis-based service selection [7 12ndash14] reputation-aware service selection [15] adaptive learn-ing mechanism-based service selection [8 16] economictheoretical model-based service selection [17 18] servicelevel agreement-based service ranking [6] visualizationframework for service selection [19] and trust evaluationmiddleware for cloud service selection [20] Though theseapproaches can efficiently measure service quality the imple-mentation of some approaches is time-consuming and costly

Ma et al [7] propose a time-aware service selectionapproach by using interval neutrosophic set In the paper thestrategy for selecting trustworthy services from an abundantfield of candidates involves formulating the problem of time-aware service selection with tradeoffs between performance-costs and potential risk as a multicriterion decision-making

(MCDM) problem that creates a ranked services list usinginterval neutrosophic set (INS) theory The experimentalresults demonstrate that the proposed approach can workeffectively in both the risk-sensitive service selection modeand the performance-cost-sensitive service selection modebut the approach ignores the changes of user context infor-mation in MCC environments

Whaiduzzaman et al [12] identify and synthesize severalmulticriteria decision analysis (MCDA) techniques and pro-vide a comprehensive analysis of this technology for generalreaders In addition this paper presents a taxonomy derivedfrom a survey of the current literature The results show thatMCDA techniques are indeed effective and can be used forcloud service selection but that different techniques do notselect the same service

This paper [14] presents a cloud service selectionmethod-ology that utilizesQuality-of-Service history of cloud servicesover different time periods and performs parallel multicri-teria decision analysis to rank all cloud services in eachtime period in accordance with user preferences beforeaggregating the results to determine the overall rank ofall the available options for cloud service selection Thismethodology assists the cloud service user to select the bestpossible available service according to the requirements Themain disadvantages are that the framework proposed in thispaper deals with service selection in the preinteraction periodonly Work on postinteraction service migration decisions isneeded and several other important factors such as the costof migration in terms of service disruption and data transferalso need to be included in the decision-making process

Fan et al [15] propose a multidimensional trust-awarecloud service selection mechanism based on evidential rea-soning approach which aggregates multidimensional trustfeedback ratings to form the reputation values of the cloudservice providersThough the experimental results show thatthis approach is effective in cloud systems it can be only usedto recommend an optimal cloud service which satisfies thekey QoS requirements for users

Wang et al [16] propose a dynamic cloud service selectionmethod by using an adaptive learning mechanism whichinvolves incentive forgetting and degenerate functions thatcan realize the self-adaptive regulation for optimizing nextservice selection according to the status of current serviceselection To assist users to efficiently select their preferredcloud services a cloud service selection model adopting thecloud service brokers is given in the paper However inthe paper the brokerage scheme on cloud service selectiontypically assumes that brokers are completely trusted and donot provide any guarantee over the correctness of the servicerecommendations It is then possible for a compromisedor dishonest broker to easily take advantage of the limitedcapabilities of the clients and provide incorrect or incompleteresponses

Do et al [17] propose a dynamic service selectionmethodwhich provides a price game in heterogeneous cloud marketThe paper studies price competition in a heterogeneous cloudmarket where users can identify benefit and cost of cloudservice application and choose the best one through ana-lyzing market-relevant factors But this paper only considers

Wireless Communications and Mobile Computing 5

one service In real-world applications there are many cloudservices in the practical cloudmarket Additionally this paperhas not considered service level agreement issues that are alsoimportant for cloud users

The visualization framework for service selection [19]takes into cognizance the set of cloud services that matchesa userrsquos request and based on QoS attributes users caninteract with the results via bubble graph visualization tocompare and contrast the search results to ascertain the bestalternative Although the result from the experiments showsthat visualization framework simplifies decision-making theuse of bubble graph introduces additional complexity for theuser when making a suitable service selection

The authors [20] discuss themethod of enhancing servicetrust evaluation and propose a trustworthy selection frame-work for cloud service selection named TRUSS Aimingat developing an effective trust evaluation middleware forTRUSS this paper proposes an integrated trust evalua-tion method via combining objective trust assessment andsubjective trust assessment Simulation-based experimentsvalidated the performance of the proposed method but themethod assumes that the majority of service users are honestand a dishonest user gives more unfair ratings than fairratings This assumption is unrealistic

Cloud computing and mobile cloud computing are quitedifferentThemobile cloudpays attention to services availablethrough mobile network operators (MNOs like Verizon andATampT) While a great number of researchers have focusedon the service selection and recommendation in cloudcomputing little attention has been devoted to trustworthyservice selection in mobile cloud computing Different fromthese existing approaches our work focuses on how to selecttrustworthy and high quality cloud services formobile user inMCC which is an urgently required research problemThuswe propose a context-aware cloud service selection methodfor mobile cloud computing environments which considersuser context information and interest in order to calculateuser similarities and utilize relational degree among cloudservices to improve the accuracy of service ranking results

3 SSRM

In this section we present a detailed discussion of theproposed novel service selection and recommendationmodel(SSRM) Firstly user similarities are calculated and thebasic personalized service ranking results are obtained inSection 31 Secondly relational degree is calculated and usedto improve the accuracy of ranking results in Section 32

31 User Similarity Computation Given119898users and 119899mobilecloud services the user-service matrix (USM) for predictingthe missing service usage experiences is denoted as

[[[[1199031199061 mcs1 sdot sdot sdot 1199031199061 mcs119899 d

119903119906119898 mcs1 sdot sdot sdot 119903119906119898 mcs119899

]]]] (1)

where 119903119906119898 mcs119899 expresses historical service usage experiences(QoS value or customer rating) of mobile cloud service mcs119899made by user 119906119898 ldquo119903119898119899 = nullrdquo states that 119906119898 did notinvoke mcs119899 yet In order to get the ranking results we firstlycalculate user similarityThe user similarity is measured withthe following equation

sim119880 (119906119909 119906119910) = 120582 lowast sim119862 (119906119909 119906119910)+ (1 minus 120582) sim119868 (119906119909 119906119910) (2)

where 119906119909 and 119906119910 denote two users sim119862(119906119909 119906119910) andsim119868(119906119909 119906119910) respectively express context similarity and userinterest similarity between 119906119909 and 119906119910 120582 is defined to deter-mine how much similarity measure relies on context andinterest 120582 is in the interval [0 1]311 Context Similarity Computation Cloud service selec-tion of a user in a context (eg a laptop with enough battery)may bemuch different from that of another user in a differentcontext (eg a smart phone without enough battery) Hencecloud service selection can be significantly affected by usercontext To evaluate the context similarities between targetuser and training user we provide the definition of contextThe definition from [21] is referenced by us Let 119903119906119909 mcs119895 beusage experience of cloud service mcs119895 made by user 119906119909Definition 1 (context) Context 119862119906119909 mcs119895 is any informationthat can be used to characterize the situation where usageexperience 119903119906119909 mcs119895 of cloud service mcs119895 is made by user 119906119909

Context information has different properties whichinclude location resource ability and status Let V be the typeof context property of 119862119906119909 mcs119895

The context 119862119906119909 mcs119895 is denoted as

119862119906119909 mcs119895 = (119862nor119906119909mcs119895 (1) 119862nor

119906119909mcs119895 (2) 119862nor119906119909mcs119895 (V)) (3)

where 119862nor119906119909mcs119895(V) is the normalized property value

Pearson Correlation Coefficient (PCC) has been success-fully employed to obtain the numerical distance betweendifferent users for similarity calculation [4 5] Let 119906119909 and 119906119910be two users then PCC is applied to calculate the contextsimilarity between 119906119909 and 119906119910 over service mcs119895 by

sim119862mcs119895(119906119909 119906119910) = sumV

119904=1 (119862nor119906119909 mcs119895 (119904) minus 119862nor

119906119909mcs119895 (119904)) (119862nor119906119910mcs119895 (119904) minus 119862nor

119906119910mcs119895 (119904))radicsumV119904=1 (119862nor

119906119909mcs119895 (119904) minus 119862nor119906119909mcs119895 (119904))2radicsumV

119904=1 (119862nor119906119910mcs119895 (119904) minus 119862nor

119906119910 mcs119895 (119904))2 (4)

6 Wireless Communications and Mobile Computing

where sim119862mcs119895(119906119909 119906119910) expresses the context similarity

between 119906119909 and 119906119910 and it is in the interval of [minus1 1] and119862119906119909 mcs119895(119904) and 119862119906119910 mcs119895(119904) stand for the average values ofcontext

312 Interest Similarity Computation In mobile cloud com-puting environments each user lives in a large network offriends which is called social network It has been reportedthat user interestsrsquo similarity can be leveraged to supportservices and products recommendation [22] A user prefersto choose the items recommended by other users withsimilar interest in a social network Exploring those usersof a high interest similarity with the existing clients couldefficiently enlarge client groups for cloud service providers[22]

Here we use cosine similarity to compute interest similar-ity similar to [22] Interest similarity between users 119906119909 and 119906119910is then defined as the cosine distance between their respectivecloud service invocation sets

sim119868 (119906119909 119906119910) = 100381710038171003817100381710038171003817119868119906119909 cap 1198681199061199101003817100381710038171003817100381710038171003817100381710038171003817100381711986811990611990910038171003817100381710038171003817 sdot 100381710038171003817100381710038171003817119868119906119910100381710038171003817100381710038171003817 (5)

where 119868119906119909 = radic119897119906119909 (119897119906119909 is the number ofmobile cloud serviceswhich have been invoked by 119906119909) and 119868119906119909 cap119868119906119910 is the numberof mobile cloud services which have been coinvoked by 119906119909and 119906119910 If 119897119906119909 = 0 or 119897119906119910 = 0 sim119868(119906119909 119906119910) is undefined

Cloud service invocation is an important factor to deter-mine the interest of users [5] For example user 119906119909 hasinvoked mobile cloud services mcs1 mcs3 mcs6 and mcs7119906119910 has invoked mcs6 and mcs7 and 119906119911 has invoked mcs1 andmcs2 Though 119906119909 119906119910 and 119906119911 do not know each other in realworld and 119906119909 and 119906119910 have a high interest similarity than 119906119909and 119906119911 as they both invoked mcs6 and mcs7 [4]

313 Getting Initial Ranking Results We first predict themissing value 119903119906119909 mcs119895 of mobile cloud service mcs119895 madeby user 119906119898 before making initial ranking results For everymobile cloud service only the Top-119896 similar training useris selected to make missing value prediction Based on thesim119880(119906119909 119906119910) values a set of the Top-k similar training users119873119862mcs119895

(119906119909) over service mcs119895 is identified for target user 119906119909by

119873119862mcs119895(119906119909)

= 119906119910 | 119906119910 isin TS119906119909 sim119880 (119906119909 119906119910) gt 0 119906119909 = 119906119910 (6)

where TS119906119909 is a set of the Top-k similar training users to targetuser 119906119909 and sim119880(119906119909 119906119910) gt 0 excludes the dissimilar userwith negative similar values The value of sim119880(119906119909 119906119910) in (6)is calculated by (2) The missing value 119903119906119909mcs119895 is predicted asfollows

119903119906119909mcs119895= 119903119906119909+ sum119906119910isin119873119862mcs119895

(119906119909)(119903119906119910 mcs119895 minus 119903119906119910) sim119880 (119906119909 119906119910)sum119906119910isin119873119862mcs119895(119906119909)

sim119880 (119906119909 119906119910) (7)

where 119903119906119909 and 119903119906119910 are the average value of service usageexperiences (QoS value or customer rating) of mobile cloudservice made by users 119906119909 and 119906119910 respectively As differentQoS properties have different dimensions and range of valueswe first ensure predicted missing values in the range of [0 1]In [5] QoS properties are classified into two categories ldquocostrdquoand ldquobenefitrdquo For ldquocostrdquo property (response-time) the lowerits value is the greater possibility that a user would chooseit becomes In SSRM all QoS properties are consideredas ldquobenefitrdquo attribute Normalized missing value 119876119906119909 mcs119895 iscomputed as follows

119876119906119909 mcs119895

=

119903119906119909mcs119895 minus min (119903119906119909)max (119903119906119909) minus min (119903119906119909) 119903119906119909mcs119895 isin ldquobenefitrdquo

max (119903119906119909) minus 119903119906119909mcs119895

max (119903119906119909) minus min (119903119906119909) 119903119906119909mcs119895 isin ldquocostrdquo(8)

where min(119903119906119909) and max(119903119906119909) denote the minimum andmaximumQoSproperty value for user119906119909 and they are subjectto the following constraints

min (119903119906119909) = min (119903119906119909mcs119895 | 119895 = 1 119899)max (119903119906119909) = max (119903119906119909 mcs119895 | 119895 = 1 119899) (9)

119876119906119909 mcs119895 is in the range of [0 1] The larger its value isthe more possibility that user 119906119909 would be satisfied with theservice becomes Finally the cloud services are ranked for 119906119909in the order of decreasing 119876119906119909mcs119895 values for cloud serviceselection similar to [4]The basic ranking algorithm is shownin Algorithm 1 The basic ranking algorithm is similar to [3]Compared with [3ndash5] in our ranking algorithm the usersimilarity is measured based on user context informationand user interest In addition the basic ranking algorithm isenhanced in Section 32 In the enhanced ranking algorithmwe calculate the relational degree among cloud services andutilize it to improve the accuracy of ranking results

The basic algorithm ranks the employed mobile cloudservice in 119864 based on the observed historical service usageexperiences The ranking results are stored in 119877119890(119905) which isin the interval [1 |119864|] A smaller value shows a higher rank119905 expresses a mobile cloud service For every cloud servicemcs119895 normalized value 119876119906119909 mcs119895 is calculated Services areranked from high to low by picking up the service 119905 that hasthe maximum 119876119906119909mcs119895 value

Wireless Communications and Mobile Computing 7

Inputan employed service set 119864a full service set 119868an employed similar training user set 119873119862mcs119895

(119906119909)observed service usage experience values 119903119906119910 mcs119895 (119906119910 isin 119873119862mcs119895

(119906119909))Outa service ranking 119865 = 119864while 119865 = Oslash do119905 = argmaxmcs119895isin119865119876119906119909 mcs119895 119877119890(119905) = |119864| minus |119865| + 1119865 = 119865 minus 119905endforeach mcs119895 isin 119868 do

119876119906119909 mcs119895 =

119903119906119909 mcs119895 minus min(119903119906119909 )max(119903119906119909 ) minus min(119903119906119909 ) 119903119906119909 mcs119895 isin ldquobenefitrdquo

max(119903119906119909 ) minus 119903119906119909 mcs119895

max(119903119906119909 ) minus min(119903119906119909 ) 119903119906119909 mcs119895 isin ldquocostrdquo

end119899 = |119868|while 119868 = Oslash do119905 = argmaxmcs119895isin119868119876119906119909 mcs119895(119905) = 119899 minus |119868| + 1119868 = 119868 minus 119905end

end

Algorithm 1 Basic ranking algorithm

32 Improving Accuracy of Ranking Results Through calcu-lating the similar relation among users the missing value119903119906119909mcs119895 is successfully predicted in Section 31 However therelation among cloud services is no doubt an importantevaluation factor of missing value prediction The relatedservices will be probably selected by the same user based onthe intuition [9]Therefore we calculate the relational degreeand utilize it to improve the accuracy of missing value 119903119906119909mcs119895computation in the section The relational degree betweenservice mcs119895 and service mcs119894 is denoted as 119889119895119894 In the paperPropFlow [23] is used to calculate 119889119895119894 PropFlow algorithmcomputes the information flow between services where alarger value indicates tighter relation In order to calculate119889119895119894 we firstly describe the formal definition of service relationgraph

Definition 2 (service relation graph) Given a set of servicesMCS and a totally ordered domain of weights 119882 a servicerelation graph is a weighted undirected graph 119866 = (MCS 119864)The edge (mcs119894mcs119895 119908119894119895) in the set 119864 isin 119880 times 119880 times 119882 encodesthe link weight119908119894119895 (119908119894119895 ge 0) between service mcs119894 and servicemcs119895 An edge depicts the direct link relationship betweenmcs119894 and mcs119895 As 119866 is undirected 119908119894119895 = 119908119895119894

Equation (10) shows how to calculate 119889119895119894119889119895119894 = Input119895 lowast 119908119895119894sum119896isin119873(119895) 119908119895119896 (10)

Ｇ＝Ｍℎ

Ｇ＝ＭiＧ＝Ｍl

Ｇ＝Ｍe

Ｇ＝Ｍf

Ｇ＝Ｍj Ｇ＝Ｍk

wkl = 1 wli = 5

wle = 1

wie = 1wke = 1

wkℎ = 2

wjk = 3

wjf = 1

Figure 3 An example of service relation graph

where initial input is regarded as 1 and 119873(119895) is the set ofneighbors of mcs119895

Figure 3 shows an example of relation degree computa-tion where there are six kinds of mobile cloud services mcs119894and mcs119895 are indirectly linked We assume mcs119895 is startingnode

Four paths can reach mcs119894 from mcs119895 but only twopaths are considered for computation based on PropFlow

8 Wireless Communications and Mobile Computing

algorithm They are mcs119895 rarr mcs119896 rarr mcs119897 rarr mcs119894 andmcs119895 rarr mcs119896 rarr mcs119890 rarr mcs119894

First we compute 119889119895119896 The sum of link weights of mcs119895and its neighbor is 4 119889119895119896 is computed as

119889119895119896 = 1 lowast 3(1 + 3) = 1 lowast 34 = 34 (11)

119889119896119897 is computed as

119889119896119897 = 119889119895119896 lowast 1(1 + 1 + 2) = 34 lowast 14 = 316 (12)

With the same method 119889119895119894 is computed as

119889119895119894 = 119889119897119894 + 119889119890119894 = 119889119896119897 lowast 55 + 119889119896119890 lowast 11 = 38 (13)

More details of calculating 119889119895119894 are shown in [24]Next we assume service mcs119894 is invoked recently by user119906119909 119889119895119894 is incorporated into (7) and then the missing value119903119906119909mcs119895 is computed with (14) in basic ranking algorithm

Finally the ultimate ranking results are got

119903119906119909 mcs119895 = (119903119906119909+ sum119906119910isin119873119862mcs119895

(119906119909)(119903119906119910 mcs119895 minus 119903119906119910) sim119880 (119906119909 119906119910)sum119906119910isin119873119862mcs119895(119906119909)

sim119880 (119906119909 119906119910) )sdot (1 + 119889119895119894)

(14)

In SSRM we only select trusted mobile cloud serviceas candidates Therefore a set of trusted candidates areidentified for 119906119909 by

MCS119909 = mcs119895 | trustmcs119895 gt 120583119909 119895 = 1 119873 (15)

where trustmcs119895 denotes the trustworthiness of mobile cloudservice mcs119895 and 120583119909 expresses the trust threshold decidedby user 119906119909 Many existing methods [4ndash6] can be applied tocompute trustmcs119895 Here how to compute trustmcs119895 is not tobe discussed We omit the details for brevity

4 Experimental Study

In this section in order to evaluate the effectiveness ofSSRM a series of test scenarios are developed To studythe prediction performance we compare SSRM with othertwo traditional approaches item-based CF approach (IBCF)and user-based CF approach (UBCF) SSRM IBCF andUBCF are rating-oriented methods which rank the cloudservices based on the predicted QoS values Since there isno suitable real data supporting mobile cloud computingsimulation we use ns3 simulator to generate the experimentaldataset Firstly we generated cloud service invocation codes

by Axis2 [25] a Java-based open source package for cloudservices Then we simulate mobile cloud service on ns3[26] based on the invocation codes On ns3 the creationof servers mobile users and service model was easier thanwith other network simulators In the simulating process ofdataset the real-world web service QoS dataset from WS-DREAM team [3] is referenced by usThe detailed real-worldQoS values are publicly released online [27] which makesour experimental evaluations reproducible We simulate 100mobile users 25 servers and 300 services The value of linkweight between two services is randomly generated which isin the range of [1 5] In order to conduct our experimentsrealistically the changes of user context information aresimulated Three types of context are considered includingbandwidth memory and battery capacity The bandwidthvalue of a mobile user is in the range of 4Mndash8MThe valuesof memory and battery capacity will decline over timeThesevalues are in the range of 100000000ndash2100000000MB and05ndash2KVA

We normalize the context information by

119862nor119906119909mcs119895 (V) = 119886 tan (119862119906119909 mcs119895 (V)) lowast 2120587 (16)

where 119862nor119906119909mcs119895(V) denotes the normalized context value of

user 119906119909 over cloud service mcs119895 Inverse tangent function isapplied to normalize the raw data 119862119906119909 mcs119895(V) Then thesenormalized property values are used to calculate contextsimilarity with (4) Once the context simulation is finishedthe target user 119906119909 will obtain context similarity results foreach cloud service

Similar to [3] QoS properties include throughput andresponse-time that are in the range of 0ndash1000 kbps and 0ndash20second respectively Most of the throughput and response-time values fall between 5ndash40 kbps and 01ndash08 seconds Inorder to simply the experimental process 200 service QoSrecords are randomly selected and two 100 times 200 user-servicematrices are constructed The two matrices include through-put and response-time respectivelyUser-servicematrices areseparated into two parts training set (80 historical usageexperiences in the matrix) and test set (the remaining 20usage experiences) Each entry in the matrix is the QoSvalue (eg response-time or throughput) of a mobile cloudservice observed by a user The experiments employ the QoSvalues of response-time and throughput to rank the servicesindependently Table 1 shows descriptions of the obtainedpractical cloud service QoS values and context information

We use Mean Absolute Error (MAE) and Root MeanSquare Error (RMSE) as the metric to evaluate predictionperformance of the proposed approach in comparison withother approaches MAE and RMSE are defined as

MAE = sum119906119909 mcs119895

100381610038161003816100381610038161003816119903act119906119909mcs119895 minus 119903pre119906119909mcs119895100381610038161003816100381610038161003816119897pre

RMSE = radic sum119906119909 mcs119895 (119903act119906119909mcs119895 minus 119903pre119906119909 mcs119895)2119897pre (17)

Wireless Communications and Mobile Computing 9

Response-time

02

025

03

035

04

045

05

055

06

065

07

MA

E

01 02 03 04 05 06 07 08 09 100Weighting factor

(a) MAE values

Response-time

03

035

04

045

05

055

06

065

07

075

08

RMSE

01 02 03 04 05 06 07 08 09 100Weighting factor

(b) RMSE values

Figure 4 Impact of weighting factor 120582

Table 1 Mobile cloud service QoS dataset and context informationdescriptions

Statistics ValueNumber of mobile cloud serviceinvocations 180000

Number of service users 100Number of mobile cloud services 200Minimum response-time value 0004 sMaximum response-time value 20 sMean of response-time 110 sStandard deviation ofresponse-time 226 s

Minimum throughput value 01 kbpsMaximum throughput value 1000 kbpsMean of throughput 2546 kbpsStandard deviation of throughput 4803 kbpsBandwidth 4Mndash8MMemory 100000000MBndash2100000000MBBattery capacity 05 KVAndash2KVA

where 119903act119906119909mcs119895 and 119903pre119906119909mcs119895 denote the actual QoS valueand predicted value respectively 119897pre is the number ofpredicted QoS values Smaller values of MAE and RMSEindicate better results

41 Impact of 120582 120582 is weighting factor in (2) 120582 determineshow much similarity measure relies on context and interestIn the experiment we vary the weighting factor 120582 from 0 to 1in increment of 01 The Top-k is set to 5 Matrix density is setto 10 Matrix density means the percentage of selected QoSentries that are used to predict the missing QoS value Theexperimental results are shown in Figure 4 As the value of 120582increases MAE firstly decreases and then quickly increasesWhen 120582 = 04 our results have the best value As shown inFigure 4(b) RMSE presents similar trend

42 Performance Comparison of SSRM IBCF and UBCFTo compare the performance of SSRM IBCF and UBCFsimilarity computation methods we implement IBCF andUBCF methods

Item-Based CF Approach (IBCF) We apply PCC to calculatesimilarities between users and predicted QoS values based onsimilar users The user similarity is calculated by

sim (119906119909 119906119910) = summcs119895isinmcs119906119909119906119910(119903119906119909 mcs119895 minus 119903119906119909) (119903119906119910 mcs119895 minus 119903119906119910)

radicsummcs119895isinmcs119906119909119906119910(119903119906119909mcs119895 minus 119903119906119909)2radicsummcs119895isinmcs119906119909119906119910

(119903119906119910mcs119895 minus 119903119906119910)2 (18)

10 Wireless Communications and Mobile Computing

where mcs119906119909119906119910 is the set of mobile cloud services that havebeen coinvoked by 119906119909 and 119906119910 119903119906119909and 119903119906119910 are average QoSvalues of cloud service invoked by 119906119909 and 119906119910 respectively

User-Based CF Approach (UBCF) We apply PCC to calculatesimilarities between services and predicted QoS values basedon similar services The service similarity is calculated by

sim (mcs119895mcs119894) = sum119906119909isin119906mcs119895mcs119894(119903119906119909mcs119895 minus 119903mcs119895) (119903119906119909 mcs119894 minus 119903mcs119894)

radicsum119906119909isin119906mcs119895mcs119894(119903119906119909mcs119895 minus 119903mcs119895)2radicsum119906119909isin119906mcs119895mcs119894

(119903119906119909mcs119894 minus 119903mcs119894)2 (19)

where 119906mcs119895mcs119894 is the set of users who have invoked the cloudservices mcs119895 andmcs119894 119903mcs119895 and 119903mcs119894 are average QoS valuesof cloud services mcs119895 and mcs119894 made by 119906119909

In the experiments trustworthiness of cloud service hasno influence on similarity computation and service decisionin the experiments In order to simplify the experimentalprocess we consider all cloud services are trustworthyWeighting factor 120582 is set to 04 We firstly study the effectof neighbor size 119896 Matrix density is set to 10 We change119896 from 5 to 30 in increment of 5 Figure 5 shows theexperimental results for response-time and throughput

As shown in Figure 5 the prediction performance ofSSRM outperforms other two approaches The performanceof IBCF is similar to UBCF As the values of 119896 increase betteraccuracy can be achieved This indicates that more similarneighbor records can provide more information for missingvalue prediction However when 119896 is larger than 25 MAEand RSME fail to drop with the value of 119896 increasing Themain reason is the limited number of similar neighbors Theobservations also show that RMSE has the similar trend butwith larger fluctuations

Next we investigate the impact of matrix density Thematrix density varies from 10 to 50 in increment of10 Similar user size k is set to 5 in this experimentFigure 6 shows the experimental results for response-timeand throughput

As presented in Figure 6 the prediction performance isalso enhanced with the value of matrix density increasingThe main reason is that denser user-service matrix providesmore information for missing value prediction NARM hasthe better performance than LBCF and UBCF under allexperimental setting consistently since NARM considerscontext similarity and the relational degree among cloudservices The experimental results of LBCF and UBCF aresimilar in this experiment since these two similarity compu-tation methods are similar to each other and are both rating-oriented

43 Performance Comparison with Other Popular ApproachesTo study the prediction performance of SSRM we com-pare SSRM with three existing QoS properties predictionapproaches CloudRank2 [3] TECSS [4] and JV-PCC [5]

The size of Top-119896 similar service is an important factorin CF approach which determines how many neighborsrsquohistorical records are employed to generate predictions [5]Therefore the impact of matrix density is not considered in

the experiments We set the density to 10 We vary 119896 from 5to 30 in increment of 5 In order to simplify the experimentalprocess we consider all cloud services are trustworthyFigure 7 shows the experimental results for response-timeand throughput Under the same simulation condition SSRMsignificantly outperforms CloudRank2 TECSS and JV-PCCThe observations also suggest that better accuracy can beachieved by our model when more historical records areavailable in the service selection studyThemain reason is thatcomputing context similarity can improve the performance ofprediction evaluation in MCC environments

Figures 7(a) and 7(b) depict the MAE fractions of SSRMCloudRank2 TECSS and JV-PCC for response-time andthroughput while Figures 7(c) and 7(d) depict the RMSEfractions It can be observed that SSRM achieves smallerMAE and RMSE consistently than other approaches for bothresponse-time and throughput

5 Conclusions and Future

In the paper we propose a novel service selection and recom-mendation model (SSRM) for mobile cloud computing envi-ronmentsWhen calculating user similarity context informa-tion and interest are considered Through utilizing relationaldegree to improve the accuracy of ranking result our serviceselection and recommendation method succeed in obtainingthe ultimate service ranking results In addition SSRM onlyselects trusted mobile cloud service as candidates Thereforea set of trusted candidates are identified for target user Theexperimental results show that our approach significantlyimproves the prediction performance as comparedwith othertwo traditional approaches item-based CF approach (IBCF)and user-based CF approach (UBCF)

There are some disadvantages in our SSRM Firstly weexploit node distance to compute context similarity Thereis an underlying assumption in this exploitation each ofthe two adjacent nodes has equal semantic distance orgranularity of nodes in each level is identicalThis underlyingassumption is not true in some scenarios and it deterioratesthe performance of similarity evaluation Secondly the trustthreshold is used to select trusted mobile cloud service ascandidates However it cannot detect and exclude maliciousQoS values provided by users Thirdly in our experimentsthere are only three types of context information that areconsidered

Therefore in the future we will study the methods toimprove context similarity measurement We will study how

Wireless Communications and Mobile Computing 11

Response-time

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E10 25 3015 205

k

(b)Response-time

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

(c)

Throughput

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

(d)

Figure 5 Impact of neighbor size 119896 comparison with SSRM IBCF and UBCF

12 Wireless Communications and Mobile Computing

Response-time

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E

20 50 6030 4010Matrix density ()

(a)

Throughput

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E20 30 40 50 6010

Matrix density ()

(b)

Response-time

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 50 6030 4010Matrix density ()

(c)

Throughput

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 30 40 50 6010Matrix density ()

(d)

Figure 6 Impact of matrix density comparison with SSRM IBCF and UBCF

Wireless Communications and Mobile Computing 13

Response-time

JV-PCCTECSSSSRM

CloudRank2

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

02

025

03

035

04

045

05

055

06

065

07

MA

E10 15 20 25 305

k

JV-PCCTECSSSSRM

CloudRank2

(b)

Response-time

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

JV-PCCTECSSSSRM

CloudRank2

(c)

Throughput

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

JV-PCCTECSSSSRM

CloudRank2

(d)

Figure 7 Impact of neighbor size 119896 comparison with other popular approaches

14 Wireless Communications and Mobile Computing

to select themost trustworthy cloud service of certain type forthe active users based on multidimensional trust evidenceWe also will conduct more experimental investigations thatdeal with the impact of different context changes on serviceselection Moreover we will improve the ranking accuracyof our approaches by exploiting additional techniques (com-bining rating-oriented approaches and ranking-orientedapproaches matrix factorization intelligent optimizationalgorithms etc)

Conflicts of Interest

The author declares that they have no conflicts of interest

Acknowledgments

The work in this paper has been supported by NationalNatural Science Foundation of China (Program no 71501156)and China Postdoctoral Science Foundation (Program no2014M560796) and Shanxi Provincial EducationDepartment(Program no 15JK1679)

References

[1] White PaperMobile CloudComputing Solution Brief AEPONA2010

[2] httpsenwikipediaorgwikiCloudlet[3] Z Zheng X Wu Y Zhang M R Lyu and J Wang ldquoQoS

ranking prediction for cloud servicesrdquo IEEE Transactions onParallel and Distributed Systems vol 24 no 6 pp 1213ndash12222013

[4] Y Pan SDingW Fan J Li and S Yang ldquoTrust-enhanced cloudservice selection model based on QoS analysisrdquo PLoS ONE vol10 no 11 Article ID e0143448 2015

[5] S Ding C-Y Xia K-L Zhou S-L Yang and J S Shang ldquoDeci-sion support for personalized cloud service selection throughmulti-attribute trustworthiness evaluationrdquo PLoS ONE vol 9no 6 Article ID e107821 2014

[6] S K Garg S Versteeg and R Buyya ldquoA framework for rankingof cloud computing servicesrdquo Future Generation ComputerSystems vol 29 no 4 pp 1012ndash1023 2013

[7] H Ma Z Hu K Li and H Zhang ldquoToward trustworthy cloudservice selection a time-aware approach using interval neutro-sophic setrdquo Journal of Parallel and Distributed Computing vol96 pp 75ndash94 2016

[8] A Ahmed and A S Sabyasachi ldquoMobile cloud service selectionusing back propagation neural networkrdquo International Journalof Computer Applications vol 129 no 14 pp 1ndash5 2015

[9] L Guo J Ma Z-M Chen and H-R Jiang ldquoIncorporatingitem relations for social recommendationrdquo Chinese Journal ofComputers vol 37 no 1 pp 219ndash228 2014

[10] Z Yang B Wu K Zheng X Wang and L Lei ldquoA survey ofcollaborative filtering-based recommender systems for mobileinternet applicationsrdquo IEEE Access vol 4 pp 3273ndash3287 2016

[11] Markets and Markets httpwwwmarketsandmarketscomPressReleasesmobile-cloudasp

[12] M Whaiduzzaman A Gani N B Anuar M Shiraz MN Haque and I T Haque ldquoCloud service selection usingmulticriteria decision analysisrdquoTheScientificWorld Journal vol2014 Article ID 459375 10 pages 2014

[13] K Tserpes F Aisopos D Kyriazis and T Varvarigou ldquoArecommender mechanism for service selection in service-oriented environmentsrdquo Future Generation Computer Systemsvol 28 no 8 pp 1285ndash1294 2012

[14] Z U Rehman O K Hussain and F K Hussain ldquoParallelcloud service selection and ranking based on QoS historyrdquoInternational Journal of Parallel Programming vol 42 no 5 pp820ndash852 2014

[15] W-J Fan S-L YangH Perros and J Pei ldquoAmulti-dimensionaltrust-aware cloud service selection mechanism based on evi-dential reasoning approachrdquo International Journal of Automa-tion and Computing vol 12 no 2 pp 208ndash219 2015

[16] X G Wang J Cao and Y Xiang ldquoDynamic cloud serviceselection using an adaptive learning mechanism in multi-cloudcomputingrdquo The Journal of Systems and Software vol 100 pp195ndash210 2015

[17] C T Do N H Tran E-N Huh C S Hong D Niyato andZ Han ldquoDynamics of service selection and provider pricinggame in heterogeneous cloud marketrdquo Journal of Network andComputer Applications vol 69 pp 152ndash165 2015

[18] R Pal and P Hui ldquoEconomic models for cloud service marketspricing and capacity planningrdquo Theoretical Computer Sciencevol 496 pp 113ndash124 2013

[19] A Ezenwoke O Daramola and M Adigun ldquoTowards avisualization framework for service selection in cloud e-marketplacesrdquo in Proceedings of the 2017 IEEE InternationalConference on Web Services (ICWS rsquo17) pp 828ndash835 HonoluluHI USA June 2017

[20] M Tang X Dai J Liu and J Chen ldquoTowards a trust evaluationmiddleware for cloud service selectionrdquo Future GenerationComputer Systems vol 74 pp 302ndash312 2017

[21] A K Dey ldquoUnderstanding and using contextrdquo Personal andUbiquitous Computing vol 5 no 1 pp 4ndash7 2001

[22] X Han L Wang S Park A Cuevas and N Crespi ldquoAlikepeople alike interests A large-scale study on interest similarityin social networksrdquo in Proceedings of the 2014 IEEEACM Inter-national Conference on Advances in Social Networks Analysisand Mining (ASONAM rsquo14) pp 491ndash496 August 2014

[23] R N Lichtenwalter J T Lussier and N V Chawla ldquoNewperspectives and methods in link predictionrdquo in Proceedings ofthe 16th ACM SIGKDD International Conference on KnowledgeDiscovery and Data Mining (KDD rsquo10) pp 243ndash252 July 2010

[24] L Munasinghe and R Ichise ldquoLink prediction in social net-works using information flow via active linksrdquo IEICE Transac-tion on Information and Systems vol E96-D no 7 pp 1495ndash1502 2013

[25] Axis httpaxisapacheorgaxis2[26] ns3 httpswwwnsnamorgns3[27] Dataset httpwwwzibinzhengcomtpds

4 Wireless Communications and Mobile Computing

the most suitable mobile cloud service Therefore how toselect trustworthy and high quality mobile cloud servicesbecomes one of the most urgent problems A number ofworks have been carried out on cloud service selection issueincluding rating-oriented collaborative filtering methods [45] ranking-oriented collaborative filtering methods [3] andsome other methods

Rating-oriented or ranking-oriented collaborative filter-ing methods can produce QoS prediction or service rank-ing using collaborative filtering technology Rating-orientedcollaborative filtering methods rank the cloud services basedon the predicted QoS values QoS presents the nonfunctionalproperties of cloud services including security availabilitythroughput and response-time properties Based on theservice QoS measures various approaches are proposed forservice selection Pan et al [4] propose a trust-enhancedcloud service selection model based on QoS analysis In themodel trust is utilized to find similar neighbors and predictthemissingQoS values Ding et al [5] propose a personalizedcloud service selection method to make experience usabilityand value distribution to measure the service similarityRating-oriented collaborative filtering approaches first pre-dict the missing QoS values before making QoS rankingThe target of rating-oriented approaches is to predict QoSvalues as accurate as possible However accurate QoS valueprediction may not lead to accurate QoS ranking prediction[3]

Compared with rating-oriented methods ranking-oriented methods predict the QoS rankings directlyDifferent from the previous ranking-oriented methodsZheng et al [3] propose a comprehensive study of how toprovide accurate QoS ranking for cloud services Althoughranking-oriented methods can be used to make optimalcloud service selection from a set of functionally equivalentservice candidates these methods ignore the changes ofQoS QoS in MCCmight vary largely even for the same typeof mobile cloud services QoS in MCC are affected by theuser context information The rating-oriented collaborativefilteringmethods and ranking-oriented collaborative filteringmethods both can help user to predict missing QoS valuebut they did not consider the dynamic QoS properties inMCC

Other types of service selection approaches are alsowidely examined The most employed approaches includemulticriteria decision analysis-based service selection [7 12ndash14] reputation-aware service selection [15] adaptive learn-ing mechanism-based service selection [8 16] economictheoretical model-based service selection [17 18] servicelevel agreement-based service ranking [6] visualizationframework for service selection [19] and trust evaluationmiddleware for cloud service selection [20] Though theseapproaches can efficiently measure service quality the imple-mentation of some approaches is time-consuming and costly

Ma et al [7] propose a time-aware service selectionapproach by using interval neutrosophic set In the paper thestrategy for selecting trustworthy services from an abundantfield of candidates involves formulating the problem of time-aware service selection with tradeoffs between performance-costs and potential risk as a multicriterion decision-making

(MCDM) problem that creates a ranked services list usinginterval neutrosophic set (INS) theory The experimentalresults demonstrate that the proposed approach can workeffectively in both the risk-sensitive service selection modeand the performance-cost-sensitive service selection modebut the approach ignores the changes of user context infor-mation in MCC environments

Whaiduzzaman et al [12] identify and synthesize severalmulticriteria decision analysis (MCDA) techniques and pro-vide a comprehensive analysis of this technology for generalreaders In addition this paper presents a taxonomy derivedfrom a survey of the current literature The results show thatMCDA techniques are indeed effective and can be used forcloud service selection but that different techniques do notselect the same service

This paper [14] presents a cloud service selectionmethod-ology that utilizesQuality-of-Service history of cloud servicesover different time periods and performs parallel multicri-teria decision analysis to rank all cloud services in eachtime period in accordance with user preferences beforeaggregating the results to determine the overall rank ofall the available options for cloud service selection Thismethodology assists the cloud service user to select the bestpossible available service according to the requirements Themain disadvantages are that the framework proposed in thispaper deals with service selection in the preinteraction periodonly Work on postinteraction service migration decisions isneeded and several other important factors such as the costof migration in terms of service disruption and data transferalso need to be included in the decision-making process

Fan et al [15] propose a multidimensional trust-awarecloud service selection mechanism based on evidential rea-soning approach which aggregates multidimensional trustfeedback ratings to form the reputation values of the cloudservice providersThough the experimental results show thatthis approach is effective in cloud systems it can be only usedto recommend an optimal cloud service which satisfies thekey QoS requirements for users

Wang et al [16] propose a dynamic cloud service selectionmethod by using an adaptive learning mechanism whichinvolves incentive forgetting and degenerate functions thatcan realize the self-adaptive regulation for optimizing nextservice selection according to the status of current serviceselection To assist users to efficiently select their preferredcloud services a cloud service selection model adopting thecloud service brokers is given in the paper However inthe paper the brokerage scheme on cloud service selectiontypically assumes that brokers are completely trusted and donot provide any guarantee over the correctness of the servicerecommendations It is then possible for a compromisedor dishonest broker to easily take advantage of the limitedcapabilities of the clients and provide incorrect or incompleteresponses

Do et al [17] propose a dynamic service selectionmethodwhich provides a price game in heterogeneous cloud marketThe paper studies price competition in a heterogeneous cloudmarket where users can identify benefit and cost of cloudservice application and choose the best one through ana-lyzing market-relevant factors But this paper only considers

Wireless Communications and Mobile Computing 5

one service In real-world applications there are many cloudservices in the practical cloudmarket Additionally this paperhas not considered service level agreement issues that are alsoimportant for cloud users

The visualization framework for service selection [19]takes into cognizance the set of cloud services that matchesa userrsquos request and based on QoS attributes users caninteract with the results via bubble graph visualization tocompare and contrast the search results to ascertain the bestalternative Although the result from the experiments showsthat visualization framework simplifies decision-making theuse of bubble graph introduces additional complexity for theuser when making a suitable service selection

The authors [20] discuss themethod of enhancing servicetrust evaluation and propose a trustworthy selection frame-work for cloud service selection named TRUSS Aimingat developing an effective trust evaluation middleware forTRUSS this paper proposes an integrated trust evalua-tion method via combining objective trust assessment andsubjective trust assessment Simulation-based experimentsvalidated the performance of the proposed method but themethod assumes that the majority of service users are honestand a dishonest user gives more unfair ratings than fairratings This assumption is unrealistic

Cloud computing and mobile cloud computing are quitedifferentThemobile cloudpays attention to services availablethrough mobile network operators (MNOs like Verizon andATampT) While a great number of researchers have focusedon the service selection and recommendation in cloudcomputing little attention has been devoted to trustworthyservice selection in mobile cloud computing Different fromthese existing approaches our work focuses on how to selecttrustworthy and high quality cloud services formobile user inMCC which is an urgently required research problemThuswe propose a context-aware cloud service selection methodfor mobile cloud computing environments which considersuser context information and interest in order to calculateuser similarities and utilize relational degree among cloudservices to improve the accuracy of service ranking results

3 SSRM

In this section we present a detailed discussion of theproposed novel service selection and recommendationmodel(SSRM) Firstly user similarities are calculated and thebasic personalized service ranking results are obtained inSection 31 Secondly relational degree is calculated and usedto improve the accuracy of ranking results in Section 32

31 User Similarity Computation Given119898users and 119899mobilecloud services the user-service matrix (USM) for predictingthe missing service usage experiences is denoted as

[[[[1199031199061 mcs1 sdot sdot sdot 1199031199061 mcs119899 d

119903119906119898 mcs1 sdot sdot sdot 119903119906119898 mcs119899

]]]] (1)

where 119903119906119898 mcs119899 expresses historical service usage experiences(QoS value or customer rating) of mobile cloud service mcs119899made by user 119906119898 ldquo119903119898119899 = nullrdquo states that 119906119898 did notinvoke mcs119899 yet In order to get the ranking results we firstlycalculate user similarityThe user similarity is measured withthe following equation

sim119880 (119906119909 119906119910) = 120582 lowast sim119862 (119906119909 119906119910)+ (1 minus 120582) sim119868 (119906119909 119906119910) (2)

where 119906119909 and 119906119910 denote two users sim119862(119906119909 119906119910) andsim119868(119906119909 119906119910) respectively express context similarity and userinterest similarity between 119906119909 and 119906119910 120582 is defined to deter-mine how much similarity measure relies on context andinterest 120582 is in the interval [0 1]311 Context Similarity Computation Cloud service selec-tion of a user in a context (eg a laptop with enough battery)may bemuch different from that of another user in a differentcontext (eg a smart phone without enough battery) Hencecloud service selection can be significantly affected by usercontext To evaluate the context similarities between targetuser and training user we provide the definition of contextThe definition from [21] is referenced by us Let 119903119906119909 mcs119895 beusage experience of cloud service mcs119895 made by user 119906119909Definition 1 (context) Context 119862119906119909 mcs119895 is any informationthat can be used to characterize the situation where usageexperience 119903119906119909 mcs119895 of cloud service mcs119895 is made by user 119906119909

Context information has different properties whichinclude location resource ability and status Let V be the typeof context property of 119862119906119909 mcs119895

The context 119862119906119909 mcs119895 is denoted as

119862119906119909 mcs119895 = (119862nor119906119909mcs119895 (1) 119862nor

119906119909mcs119895 (2) 119862nor119906119909mcs119895 (V)) (3)

where 119862nor119906119909mcs119895(V) is the normalized property value

Pearson Correlation Coefficient (PCC) has been success-fully employed to obtain the numerical distance betweendifferent users for similarity calculation [4 5] Let 119906119909 and 119906119910be two users then PCC is applied to calculate the contextsimilarity between 119906119909 and 119906119910 over service mcs119895 by

sim119862mcs119895(119906119909 119906119910) = sumV

119904=1 (119862nor119906119909 mcs119895 (119904) minus 119862nor

119906119909mcs119895 (119904)) (119862nor119906119910mcs119895 (119904) minus 119862nor

119906119910mcs119895 (119904))radicsumV119904=1 (119862nor

119906119909mcs119895 (119904) minus 119862nor119906119909mcs119895 (119904))2radicsumV

119904=1 (119862nor119906119910mcs119895 (119904) minus 119862nor

119906119910 mcs119895 (119904))2 (4)

6 Wireless Communications and Mobile Computing

where sim119862mcs119895(119906119909 119906119910) expresses the context similarity

between 119906119909 and 119906119910 and it is in the interval of [minus1 1] and119862119906119909 mcs119895(119904) and 119862119906119910 mcs119895(119904) stand for the average values ofcontext

312 Interest Similarity Computation In mobile cloud com-puting environments each user lives in a large network offriends which is called social network It has been reportedthat user interestsrsquo similarity can be leveraged to supportservices and products recommendation [22] A user prefersto choose the items recommended by other users withsimilar interest in a social network Exploring those usersof a high interest similarity with the existing clients couldefficiently enlarge client groups for cloud service providers[22]

Here we use cosine similarity to compute interest similar-ity similar to [22] Interest similarity between users 119906119909 and 119906119910is then defined as the cosine distance between their respectivecloud service invocation sets

sim119868 (119906119909 119906119910) = 100381710038171003817100381710038171003817119868119906119909 cap 1198681199061199101003817100381710038171003817100381710038171003817100381710038171003817100381711986811990611990910038171003817100381710038171003817 sdot 100381710038171003817100381710038171003817119868119906119910100381710038171003817100381710038171003817 (5)

where 119868119906119909 = radic119897119906119909 (119897119906119909 is the number ofmobile cloud serviceswhich have been invoked by 119906119909) and 119868119906119909 cap119868119906119910 is the numberof mobile cloud services which have been coinvoked by 119906119909and 119906119910 If 119897119906119909 = 0 or 119897119906119910 = 0 sim119868(119906119909 119906119910) is undefined

Cloud service invocation is an important factor to deter-mine the interest of users [5] For example user 119906119909 hasinvoked mobile cloud services mcs1 mcs3 mcs6 and mcs7119906119910 has invoked mcs6 and mcs7 and 119906119911 has invoked mcs1 andmcs2 Though 119906119909 119906119910 and 119906119911 do not know each other in realworld and 119906119909 and 119906119910 have a high interest similarity than 119906119909and 119906119911 as they both invoked mcs6 and mcs7 [4]

313 Getting Initial Ranking Results We first predict themissing value 119903119906119909 mcs119895 of mobile cloud service mcs119895 madeby user 119906119898 before making initial ranking results For everymobile cloud service only the Top-119896 similar training useris selected to make missing value prediction Based on thesim119880(119906119909 119906119910) values a set of the Top-k similar training users119873119862mcs119895

(119906119909) over service mcs119895 is identified for target user 119906119909by

119873119862mcs119895(119906119909)

= 119906119910 | 119906119910 isin TS119906119909 sim119880 (119906119909 119906119910) gt 0 119906119909 = 119906119910 (6)

where TS119906119909 is a set of the Top-k similar training users to targetuser 119906119909 and sim119880(119906119909 119906119910) gt 0 excludes the dissimilar userwith negative similar values The value of sim119880(119906119909 119906119910) in (6)is calculated by (2) The missing value 119903119906119909mcs119895 is predicted asfollows

119903119906119909mcs119895= 119903119906119909+ sum119906119910isin119873119862mcs119895

(119906119909)(119903119906119910 mcs119895 minus 119903119906119910) sim119880 (119906119909 119906119910)sum119906119910isin119873119862mcs119895(119906119909)

sim119880 (119906119909 119906119910) (7)

where 119903119906119909 and 119903119906119910 are the average value of service usageexperiences (QoS value or customer rating) of mobile cloudservice made by users 119906119909 and 119906119910 respectively As differentQoS properties have different dimensions and range of valueswe first ensure predicted missing values in the range of [0 1]In [5] QoS properties are classified into two categories ldquocostrdquoand ldquobenefitrdquo For ldquocostrdquo property (response-time) the lowerits value is the greater possibility that a user would chooseit becomes In SSRM all QoS properties are consideredas ldquobenefitrdquo attribute Normalized missing value 119876119906119909 mcs119895 iscomputed as follows

119876119906119909 mcs119895

=

119903119906119909mcs119895 minus min (119903119906119909)max (119903119906119909) minus min (119903119906119909) 119903119906119909mcs119895 isin ldquobenefitrdquo

max (119903119906119909) minus 119903119906119909mcs119895

max (119903119906119909) minus min (119903119906119909) 119903119906119909mcs119895 isin ldquocostrdquo(8)

where min(119903119906119909) and max(119903119906119909) denote the minimum andmaximumQoSproperty value for user119906119909 and they are subjectto the following constraints

min (119903119906119909) = min (119903119906119909mcs119895 | 119895 = 1 119899)max (119903119906119909) = max (119903119906119909 mcs119895 | 119895 = 1 119899) (9)

119876119906119909 mcs119895 is in the range of [0 1] The larger its value isthe more possibility that user 119906119909 would be satisfied with theservice becomes Finally the cloud services are ranked for 119906119909in the order of decreasing 119876119906119909mcs119895 values for cloud serviceselection similar to [4]The basic ranking algorithm is shownin Algorithm 1 The basic ranking algorithm is similar to [3]Compared with [3ndash5] in our ranking algorithm the usersimilarity is measured based on user context informationand user interest In addition the basic ranking algorithm isenhanced in Section 32 In the enhanced ranking algorithmwe calculate the relational degree among cloud services andutilize it to improve the accuracy of ranking results

The basic algorithm ranks the employed mobile cloudservice in 119864 based on the observed historical service usageexperiences The ranking results are stored in 119877119890(119905) which isin the interval [1 |119864|] A smaller value shows a higher rank119905 expresses a mobile cloud service For every cloud servicemcs119895 normalized value 119876119906119909 mcs119895 is calculated Services areranked from high to low by picking up the service 119905 that hasthe maximum 119876119906119909mcs119895 value

Wireless Communications and Mobile Computing 7

Inputan employed service set 119864a full service set 119868an employed similar training user set 119873119862mcs119895

(119906119909)observed service usage experience values 119903119906119910 mcs119895 (119906119910 isin 119873119862mcs119895

(119906119909))Outa service ranking 119865 = 119864while 119865 = Oslash do119905 = argmaxmcs119895isin119865119876119906119909 mcs119895 119877119890(119905) = |119864| minus |119865| + 1119865 = 119865 minus 119905endforeach mcs119895 isin 119868 do

119876119906119909 mcs119895 =

119903119906119909 mcs119895 minus min(119903119906119909 )max(119903119906119909 ) minus min(119903119906119909 ) 119903119906119909 mcs119895 isin ldquobenefitrdquo

max(119903119906119909 ) minus 119903119906119909 mcs119895

max(119903119906119909 ) minus min(119903119906119909 ) 119903119906119909 mcs119895 isin ldquocostrdquo

end119899 = |119868|while 119868 = Oslash do119905 = argmaxmcs119895isin119868119876119906119909 mcs119895(119905) = 119899 minus |119868| + 1119868 = 119868 minus 119905end

end

Algorithm 1 Basic ranking algorithm

32 Improving Accuracy of Ranking Results Through calcu-lating the similar relation among users the missing value119903119906119909mcs119895 is successfully predicted in Section 31 However therelation among cloud services is no doubt an importantevaluation factor of missing value prediction The relatedservices will be probably selected by the same user based onthe intuition [9]Therefore we calculate the relational degreeand utilize it to improve the accuracy of missing value 119903119906119909mcs119895computation in the section The relational degree betweenservice mcs119895 and service mcs119894 is denoted as 119889119895119894 In the paperPropFlow [23] is used to calculate 119889119895119894 PropFlow algorithmcomputes the information flow between services where alarger value indicates tighter relation In order to calculate119889119895119894 we firstly describe the formal definition of service relationgraph

Definition 2 (service relation graph) Given a set of servicesMCS and a totally ordered domain of weights 119882 a servicerelation graph is a weighted undirected graph 119866 = (MCS 119864)The edge (mcs119894mcs119895 119908119894119895) in the set 119864 isin 119880 times 119880 times 119882 encodesthe link weight119908119894119895 (119908119894119895 ge 0) between service mcs119894 and servicemcs119895 An edge depicts the direct link relationship betweenmcs119894 and mcs119895 As 119866 is undirected 119908119894119895 = 119908119895119894

Equation (10) shows how to calculate 119889119895119894119889119895119894 = Input119895 lowast 119908119895119894sum119896isin119873(119895) 119908119895119896 (10)

Ｇ＝Ｍℎ

Ｇ＝ＭiＧ＝Ｍl

Ｇ＝Ｍe

Ｇ＝Ｍf

Ｇ＝Ｍj Ｇ＝Ｍk

wkl = 1 wli = 5

wle = 1

wie = 1wke = 1

wkℎ = 2

wjk = 3

wjf = 1

Figure 3 An example of service relation graph

where initial input is regarded as 1 and 119873(119895) is the set ofneighbors of mcs119895

Figure 3 shows an example of relation degree computa-tion where there are six kinds of mobile cloud services mcs119894and mcs119895 are indirectly linked We assume mcs119895 is startingnode

Four paths can reach mcs119894 from mcs119895 but only twopaths are considered for computation based on PropFlow

8 Wireless Communications and Mobile Computing

algorithm They are mcs119895 rarr mcs119896 rarr mcs119897 rarr mcs119894 andmcs119895 rarr mcs119896 rarr mcs119890 rarr mcs119894

First we compute 119889119895119896 The sum of link weights of mcs119895and its neighbor is 4 119889119895119896 is computed as

119889119895119896 = 1 lowast 3(1 + 3) = 1 lowast 34 = 34 (11)

119889119896119897 is computed as

119889119896119897 = 119889119895119896 lowast 1(1 + 1 + 2) = 34 lowast 14 = 316 (12)

With the same method 119889119895119894 is computed as

119889119895119894 = 119889119897119894 + 119889119890119894 = 119889119896119897 lowast 55 + 119889119896119890 lowast 11 = 38 (13)

More details of calculating 119889119895119894 are shown in [24]Next we assume service mcs119894 is invoked recently by user119906119909 119889119895119894 is incorporated into (7) and then the missing value119903119906119909mcs119895 is computed with (14) in basic ranking algorithm

Finally the ultimate ranking results are got

119903119906119909 mcs119895 = (119903119906119909+ sum119906119910isin119873119862mcs119895

(119906119909)(119903119906119910 mcs119895 minus 119903119906119910) sim119880 (119906119909 119906119910)sum119906119910isin119873119862mcs119895(119906119909)

sim119880 (119906119909 119906119910) )sdot (1 + 119889119895119894)

(14)

In SSRM we only select trusted mobile cloud serviceas candidates Therefore a set of trusted candidates areidentified for 119906119909 by

MCS119909 = mcs119895 | trustmcs119895 gt 120583119909 119895 = 1 119873 (15)

where trustmcs119895 denotes the trustworthiness of mobile cloudservice mcs119895 and 120583119909 expresses the trust threshold decidedby user 119906119909 Many existing methods [4ndash6] can be applied tocompute trustmcs119895 Here how to compute trustmcs119895 is not tobe discussed We omit the details for brevity

4 Experimental Study

In this section in order to evaluate the effectiveness ofSSRM a series of test scenarios are developed To studythe prediction performance we compare SSRM with othertwo traditional approaches item-based CF approach (IBCF)and user-based CF approach (UBCF) SSRM IBCF andUBCF are rating-oriented methods which rank the cloudservices based on the predicted QoS values Since there isno suitable real data supporting mobile cloud computingsimulation we use ns3 simulator to generate the experimentaldataset Firstly we generated cloud service invocation codes

by Axis2 [25] a Java-based open source package for cloudservices Then we simulate mobile cloud service on ns3[26] based on the invocation codes On ns3 the creationof servers mobile users and service model was easier thanwith other network simulators In the simulating process ofdataset the real-world web service QoS dataset from WS-DREAM team [3] is referenced by usThe detailed real-worldQoS values are publicly released online [27] which makesour experimental evaluations reproducible We simulate 100mobile users 25 servers and 300 services The value of linkweight between two services is randomly generated which isin the range of [1 5] In order to conduct our experimentsrealistically the changes of user context information aresimulated Three types of context are considered includingbandwidth memory and battery capacity The bandwidthvalue of a mobile user is in the range of 4Mndash8MThe valuesof memory and battery capacity will decline over timeThesevalues are in the range of 100000000ndash2100000000MB and05ndash2KVA

We normalize the context information by

119862nor119906119909mcs119895 (V) = 119886 tan (119862119906119909 mcs119895 (V)) lowast 2120587 (16)

where 119862nor119906119909mcs119895(V) denotes the normalized context value of

user 119906119909 over cloud service mcs119895 Inverse tangent function isapplied to normalize the raw data 119862119906119909 mcs119895(V) Then thesenormalized property values are used to calculate contextsimilarity with (4) Once the context simulation is finishedthe target user 119906119909 will obtain context similarity results foreach cloud service

Similar to [3] QoS properties include throughput andresponse-time that are in the range of 0ndash1000 kbps and 0ndash20second respectively Most of the throughput and response-time values fall between 5ndash40 kbps and 01ndash08 seconds Inorder to simply the experimental process 200 service QoSrecords are randomly selected and two 100 times 200 user-servicematrices are constructed The two matrices include through-put and response-time respectivelyUser-servicematrices areseparated into two parts training set (80 historical usageexperiences in the matrix) and test set (the remaining 20usage experiences) Each entry in the matrix is the QoSvalue (eg response-time or throughput) of a mobile cloudservice observed by a user The experiments employ the QoSvalues of response-time and throughput to rank the servicesindependently Table 1 shows descriptions of the obtainedpractical cloud service QoS values and context information

We use Mean Absolute Error (MAE) and Root MeanSquare Error (RMSE) as the metric to evaluate predictionperformance of the proposed approach in comparison withother approaches MAE and RMSE are defined as

MAE = sum119906119909 mcs119895

100381610038161003816100381610038161003816119903act119906119909mcs119895 minus 119903pre119906119909mcs119895100381610038161003816100381610038161003816119897pre

RMSE = radic sum119906119909 mcs119895 (119903act119906119909mcs119895 minus 119903pre119906119909 mcs119895)2119897pre (17)

Wireless Communications and Mobile Computing 9

Response-time

02

025

03

035

04

045

05

055

06

065

07

MA

E

01 02 03 04 05 06 07 08 09 100Weighting factor

(a) MAE values

Response-time

03

035

04

045

05

055

06

065

07

075

08

RMSE

01 02 03 04 05 06 07 08 09 100Weighting factor

(b) RMSE values

Figure 4 Impact of weighting factor 120582

Table 1 Mobile cloud service QoS dataset and context informationdescriptions

Statistics ValueNumber of mobile cloud serviceinvocations 180000

Number of service users 100Number of mobile cloud services 200Minimum response-time value 0004 sMaximum response-time value 20 sMean of response-time 110 sStandard deviation ofresponse-time 226 s

Minimum throughput value 01 kbpsMaximum throughput value 1000 kbpsMean of throughput 2546 kbpsStandard deviation of throughput 4803 kbpsBandwidth 4Mndash8MMemory 100000000MBndash2100000000MBBattery capacity 05 KVAndash2KVA

where 119903act119906119909mcs119895 and 119903pre119906119909mcs119895 denote the actual QoS valueand predicted value respectively 119897pre is the number ofpredicted QoS values Smaller values of MAE and RMSEindicate better results

41 Impact of 120582 120582 is weighting factor in (2) 120582 determineshow much similarity measure relies on context and interestIn the experiment we vary the weighting factor 120582 from 0 to 1in increment of 01 The Top-k is set to 5 Matrix density is setto 10 Matrix density means the percentage of selected QoSentries that are used to predict the missing QoS value Theexperimental results are shown in Figure 4 As the value of 120582increases MAE firstly decreases and then quickly increasesWhen 120582 = 04 our results have the best value As shown inFigure 4(b) RMSE presents similar trend

42 Performance Comparison of SSRM IBCF and UBCFTo compare the performance of SSRM IBCF and UBCFsimilarity computation methods we implement IBCF andUBCF methods

Item-Based CF Approach (IBCF) We apply PCC to calculatesimilarities between users and predicted QoS values based onsimilar users The user similarity is calculated by

sim (119906119909 119906119910) = summcs119895isinmcs119906119909119906119910(119903119906119909 mcs119895 minus 119903119906119909) (119903119906119910 mcs119895 minus 119903119906119910)

radicsummcs119895isinmcs119906119909119906119910(119903119906119909mcs119895 minus 119903119906119909)2radicsummcs119895isinmcs119906119909119906119910

(119903119906119910mcs119895 minus 119903119906119910)2 (18)

10 Wireless Communications and Mobile Computing

where mcs119906119909119906119910 is the set of mobile cloud services that havebeen coinvoked by 119906119909 and 119906119910 119903119906119909and 119903119906119910 are average QoSvalues of cloud service invoked by 119906119909 and 119906119910 respectively

User-Based CF Approach (UBCF) We apply PCC to calculatesimilarities between services and predicted QoS values basedon similar services The service similarity is calculated by

sim (mcs119895mcs119894) = sum119906119909isin119906mcs119895mcs119894(119903119906119909mcs119895 minus 119903mcs119895) (119903119906119909 mcs119894 minus 119903mcs119894)

radicsum119906119909isin119906mcs119895mcs119894(119903119906119909mcs119895 minus 119903mcs119895)2radicsum119906119909isin119906mcs119895mcs119894

(119903119906119909mcs119894 minus 119903mcs119894)2 (19)

where 119906mcs119895mcs119894 is the set of users who have invoked the cloudservices mcs119895 andmcs119894 119903mcs119895 and 119903mcs119894 are average QoS valuesof cloud services mcs119895 and mcs119894 made by 119906119909

In the experiments trustworthiness of cloud service hasno influence on similarity computation and service decisionin the experiments In order to simplify the experimentalprocess we consider all cloud services are trustworthyWeighting factor 120582 is set to 04 We firstly study the effectof neighbor size 119896 Matrix density is set to 10 We change119896 from 5 to 30 in increment of 5 Figure 5 shows theexperimental results for response-time and throughput

As shown in Figure 5 the prediction performance ofSSRM outperforms other two approaches The performanceof IBCF is similar to UBCF As the values of 119896 increase betteraccuracy can be achieved This indicates that more similarneighbor records can provide more information for missingvalue prediction However when 119896 is larger than 25 MAEand RSME fail to drop with the value of 119896 increasing Themain reason is the limited number of similar neighbors Theobservations also show that RMSE has the similar trend butwith larger fluctuations

Next we investigate the impact of matrix density Thematrix density varies from 10 to 50 in increment of10 Similar user size k is set to 5 in this experimentFigure 6 shows the experimental results for response-timeand throughput

As presented in Figure 6 the prediction performance isalso enhanced with the value of matrix density increasingThe main reason is that denser user-service matrix providesmore information for missing value prediction NARM hasthe better performance than LBCF and UBCF under allexperimental setting consistently since NARM considerscontext similarity and the relational degree among cloudservices The experimental results of LBCF and UBCF aresimilar in this experiment since these two similarity compu-tation methods are similar to each other and are both rating-oriented

43 Performance Comparison with Other Popular ApproachesTo study the prediction performance of SSRM we com-pare SSRM with three existing QoS properties predictionapproaches CloudRank2 [3] TECSS [4] and JV-PCC [5]

The size of Top-119896 similar service is an important factorin CF approach which determines how many neighborsrsquohistorical records are employed to generate predictions [5]Therefore the impact of matrix density is not considered in

the experiments We set the density to 10 We vary 119896 from 5to 30 in increment of 5 In order to simplify the experimentalprocess we consider all cloud services are trustworthyFigure 7 shows the experimental results for response-timeand throughput Under the same simulation condition SSRMsignificantly outperforms CloudRank2 TECSS and JV-PCCThe observations also suggest that better accuracy can beachieved by our model when more historical records areavailable in the service selection studyThemain reason is thatcomputing context similarity can improve the performance ofprediction evaluation in MCC environments

Figures 7(a) and 7(b) depict the MAE fractions of SSRMCloudRank2 TECSS and JV-PCC for response-time andthroughput while Figures 7(c) and 7(d) depict the RMSEfractions It can be observed that SSRM achieves smallerMAE and RMSE consistently than other approaches for bothresponse-time and throughput

5 Conclusions and Future

In the paper we propose a novel service selection and recom-mendation model (SSRM) for mobile cloud computing envi-ronmentsWhen calculating user similarity context informa-tion and interest are considered Through utilizing relationaldegree to improve the accuracy of ranking result our serviceselection and recommendation method succeed in obtainingthe ultimate service ranking results In addition SSRM onlyselects trusted mobile cloud service as candidates Thereforea set of trusted candidates are identified for target user Theexperimental results show that our approach significantlyimproves the prediction performance as comparedwith othertwo traditional approaches item-based CF approach (IBCF)and user-based CF approach (UBCF)

There are some disadvantages in our SSRM Firstly weexploit node distance to compute context similarity Thereis an underlying assumption in this exploitation each ofthe two adjacent nodes has equal semantic distance orgranularity of nodes in each level is identicalThis underlyingassumption is not true in some scenarios and it deterioratesthe performance of similarity evaluation Secondly the trustthreshold is used to select trusted mobile cloud service ascandidates However it cannot detect and exclude maliciousQoS values provided by users Thirdly in our experimentsthere are only three types of context information that areconsidered

Therefore in the future we will study the methods toimprove context similarity measurement We will study how

Wireless Communications and Mobile Computing 11

Response-time

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E10 25 3015 205

k

(b)Response-time

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

(c)

Throughput

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

(d)

Figure 5 Impact of neighbor size 119896 comparison with SSRM IBCF and UBCF

12 Wireless Communications and Mobile Computing

Response-time

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E

20 50 6030 4010Matrix density ()

(a)

Throughput

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E20 30 40 50 6010

Matrix density ()

(b)

Response-time

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 50 6030 4010Matrix density ()

(c)

Throughput

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 30 40 50 6010Matrix density ()

(d)

Figure 6 Impact of matrix density comparison with SSRM IBCF and UBCF

Wireless Communications and Mobile Computing 13

Response-time

JV-PCCTECSSSSRM

CloudRank2

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

02

025

03

035

04

045

05

055

06

065

07

MA

E10 15 20 25 305

k

JV-PCCTECSSSSRM

CloudRank2

(b)

Response-time

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

JV-PCCTECSSSSRM

CloudRank2

(c)

Throughput

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

JV-PCCTECSSSSRM

CloudRank2

(d)

Figure 7 Impact of neighbor size 119896 comparison with other popular approaches

14 Wireless Communications and Mobile Computing

to select themost trustworthy cloud service of certain type forthe active users based on multidimensional trust evidenceWe also will conduct more experimental investigations thatdeal with the impact of different context changes on serviceselection Moreover we will improve the ranking accuracyof our approaches by exploiting additional techniques (com-bining rating-oriented approaches and ranking-orientedapproaches matrix factorization intelligent optimizationalgorithms etc)

Conflicts of Interest

The author declares that they have no conflicts of interest

Acknowledgments

The work in this paper has been supported by NationalNatural Science Foundation of China (Program no 71501156)and China Postdoctoral Science Foundation (Program no2014M560796) and Shanxi Provincial EducationDepartment(Program no 15JK1679)

References

[1] White PaperMobile CloudComputing Solution Brief AEPONA2010

[2] httpsenwikipediaorgwikiCloudlet[3] Z Zheng X Wu Y Zhang M R Lyu and J Wang ldquoQoS

ranking prediction for cloud servicesrdquo IEEE Transactions onParallel and Distributed Systems vol 24 no 6 pp 1213ndash12222013

[4] Y Pan SDingW Fan J Li and S Yang ldquoTrust-enhanced cloudservice selection model based on QoS analysisrdquo PLoS ONE vol10 no 11 Article ID e0143448 2015

[5] S Ding C-Y Xia K-L Zhou S-L Yang and J S Shang ldquoDeci-sion support for personalized cloud service selection throughmulti-attribute trustworthiness evaluationrdquo PLoS ONE vol 9no 6 Article ID e107821 2014

[6] S K Garg S Versteeg and R Buyya ldquoA framework for rankingof cloud computing servicesrdquo Future Generation ComputerSystems vol 29 no 4 pp 1012ndash1023 2013

[7] H Ma Z Hu K Li and H Zhang ldquoToward trustworthy cloudservice selection a time-aware approach using interval neutro-sophic setrdquo Journal of Parallel and Distributed Computing vol96 pp 75ndash94 2016

[8] A Ahmed and A S Sabyasachi ldquoMobile cloud service selectionusing back propagation neural networkrdquo International Journalof Computer Applications vol 129 no 14 pp 1ndash5 2015

[9] L Guo J Ma Z-M Chen and H-R Jiang ldquoIncorporatingitem relations for social recommendationrdquo Chinese Journal ofComputers vol 37 no 1 pp 219ndash228 2014

[10] Z Yang B Wu K Zheng X Wang and L Lei ldquoA survey ofcollaborative filtering-based recommender systems for mobileinternet applicationsrdquo IEEE Access vol 4 pp 3273ndash3287 2016

[11] Markets and Markets httpwwwmarketsandmarketscomPressReleasesmobile-cloudasp

[12] M Whaiduzzaman A Gani N B Anuar M Shiraz MN Haque and I T Haque ldquoCloud service selection usingmulticriteria decision analysisrdquoTheScientificWorld Journal vol2014 Article ID 459375 10 pages 2014

[13] K Tserpes F Aisopos D Kyriazis and T Varvarigou ldquoArecommender mechanism for service selection in service-oriented environmentsrdquo Future Generation Computer Systemsvol 28 no 8 pp 1285ndash1294 2012

[14] Z U Rehman O K Hussain and F K Hussain ldquoParallelcloud service selection and ranking based on QoS historyrdquoInternational Journal of Parallel Programming vol 42 no 5 pp820ndash852 2014

[15] W-J Fan S-L YangH Perros and J Pei ldquoAmulti-dimensionaltrust-aware cloud service selection mechanism based on evi-dential reasoning approachrdquo International Journal of Automa-tion and Computing vol 12 no 2 pp 208ndash219 2015

[16] X G Wang J Cao and Y Xiang ldquoDynamic cloud serviceselection using an adaptive learning mechanism in multi-cloudcomputingrdquo The Journal of Systems and Software vol 100 pp195ndash210 2015

[17] C T Do N H Tran E-N Huh C S Hong D Niyato andZ Han ldquoDynamics of service selection and provider pricinggame in heterogeneous cloud marketrdquo Journal of Network andComputer Applications vol 69 pp 152ndash165 2015

[18] R Pal and P Hui ldquoEconomic models for cloud service marketspricing and capacity planningrdquo Theoretical Computer Sciencevol 496 pp 113ndash124 2013

[19] A Ezenwoke O Daramola and M Adigun ldquoTowards avisualization framework for service selection in cloud e-marketplacesrdquo in Proceedings of the 2017 IEEE InternationalConference on Web Services (ICWS rsquo17) pp 828ndash835 HonoluluHI USA June 2017

[20] M Tang X Dai J Liu and J Chen ldquoTowards a trust evaluationmiddleware for cloud service selectionrdquo Future GenerationComputer Systems vol 74 pp 302ndash312 2017

[21] A K Dey ldquoUnderstanding and using contextrdquo Personal andUbiquitous Computing vol 5 no 1 pp 4ndash7 2001

[22] X Han L Wang S Park A Cuevas and N Crespi ldquoAlikepeople alike interests A large-scale study on interest similarityin social networksrdquo in Proceedings of the 2014 IEEEACM Inter-national Conference on Advances in Social Networks Analysisand Mining (ASONAM rsquo14) pp 491ndash496 August 2014

[23] R N Lichtenwalter J T Lussier and N V Chawla ldquoNewperspectives and methods in link predictionrdquo in Proceedings ofthe 16th ACM SIGKDD International Conference on KnowledgeDiscovery and Data Mining (KDD rsquo10) pp 243ndash252 July 2010

[24] L Munasinghe and R Ichise ldquoLink prediction in social net-works using information flow via active linksrdquo IEICE Transac-tion on Information and Systems vol E96-D no 7 pp 1495ndash1502 2013

[25] Axis httpaxisapacheorgaxis2[26] ns3 httpswwwnsnamorgns3[27] Dataset httpwwwzibinzhengcomtpds

Wireless Communications and Mobile Computing 5

one service In real-world applications there are many cloudservices in the practical cloudmarket Additionally this paperhas not considered service level agreement issues that are alsoimportant for cloud users

The visualization framework for service selection [19]takes into cognizance the set of cloud services that matchesa userrsquos request and based on QoS attributes users caninteract with the results via bubble graph visualization tocompare and contrast the search results to ascertain the bestalternative Although the result from the experiments showsthat visualization framework simplifies decision-making theuse of bubble graph introduces additional complexity for theuser when making a suitable service selection

The authors [20] discuss themethod of enhancing servicetrust evaluation and propose a trustworthy selection frame-work for cloud service selection named TRUSS Aimingat developing an effective trust evaluation middleware forTRUSS this paper proposes an integrated trust evalua-tion method via combining objective trust assessment andsubjective trust assessment Simulation-based experimentsvalidated the performance of the proposed method but themethod assumes that the majority of service users are honestand a dishonest user gives more unfair ratings than fairratings This assumption is unrealistic

Cloud computing and mobile cloud computing are quitedifferentThemobile cloudpays attention to services availablethrough mobile network operators (MNOs like Verizon andATampT) While a great number of researchers have focusedon the service selection and recommendation in cloudcomputing little attention has been devoted to trustworthyservice selection in mobile cloud computing Different fromthese existing approaches our work focuses on how to selecttrustworthy and high quality cloud services formobile user inMCC which is an urgently required research problemThuswe propose a context-aware cloud service selection methodfor mobile cloud computing environments which considersuser context information and interest in order to calculateuser similarities and utilize relational degree among cloudservices to improve the accuracy of service ranking results

3 SSRM

In this section we present a detailed discussion of theproposed novel service selection and recommendationmodel(SSRM) Firstly user similarities are calculated and thebasic personalized service ranking results are obtained inSection 31 Secondly relational degree is calculated and usedto improve the accuracy of ranking results in Section 32

31 User Similarity Computation Given119898users and 119899mobilecloud services the user-service matrix (USM) for predictingthe missing service usage experiences is denoted as

[[[[1199031199061 mcs1 sdot sdot sdot 1199031199061 mcs119899 d

119903119906119898 mcs1 sdot sdot sdot 119903119906119898 mcs119899

]]]] (1)

where 119903119906119898 mcs119899 expresses historical service usage experiences(QoS value or customer rating) of mobile cloud service mcs119899made by user 119906119898 ldquo119903119898119899 = nullrdquo states that 119906119898 did notinvoke mcs119899 yet In order to get the ranking results we firstlycalculate user similarityThe user similarity is measured withthe following equation

sim119880 (119906119909 119906119910) = 120582 lowast sim119862 (119906119909 119906119910)+ (1 minus 120582) sim119868 (119906119909 119906119910) (2)

where 119906119909 and 119906119910 denote two users sim119862(119906119909 119906119910) andsim119868(119906119909 119906119910) respectively express context similarity and userinterest similarity between 119906119909 and 119906119910 120582 is defined to deter-mine how much similarity measure relies on context andinterest 120582 is in the interval [0 1]311 Context Similarity Computation Cloud service selec-tion of a user in a context (eg a laptop with enough battery)may bemuch different from that of another user in a differentcontext (eg a smart phone without enough battery) Hencecloud service selection can be significantly affected by usercontext To evaluate the context similarities between targetuser and training user we provide the definition of contextThe definition from [21] is referenced by us Let 119903119906119909 mcs119895 beusage experience of cloud service mcs119895 made by user 119906119909Definition 1 (context) Context 119862119906119909 mcs119895 is any informationthat can be used to characterize the situation where usageexperience 119903119906119909 mcs119895 of cloud service mcs119895 is made by user 119906119909

Context information has different properties whichinclude location resource ability and status Let V be the typeof context property of 119862119906119909 mcs119895

The context 119862119906119909 mcs119895 is denoted as

119862119906119909 mcs119895 = (119862nor119906119909mcs119895 (1) 119862nor

119906119909mcs119895 (2) 119862nor119906119909mcs119895 (V)) (3)

where 119862nor119906119909mcs119895(V) is the normalized property value

Pearson Correlation Coefficient (PCC) has been success-fully employed to obtain the numerical distance betweendifferent users for similarity calculation [4 5] Let 119906119909 and 119906119910be two users then PCC is applied to calculate the contextsimilarity between 119906119909 and 119906119910 over service mcs119895 by

sim119862mcs119895(119906119909 119906119910) = sumV

119904=1 (119862nor119906119909 mcs119895 (119904) minus 119862nor

119906119909mcs119895 (119904)) (119862nor119906119910mcs119895 (119904) minus 119862nor

119906119910mcs119895 (119904))radicsumV119904=1 (119862nor

119906119909mcs119895 (119904) minus 119862nor119906119909mcs119895 (119904))2radicsumV

119904=1 (119862nor119906119910mcs119895 (119904) minus 119862nor

119906119910 mcs119895 (119904))2 (4)

6 Wireless Communications and Mobile Computing

where sim119862mcs119895(119906119909 119906119910) expresses the context similarity

between 119906119909 and 119906119910 and it is in the interval of [minus1 1] and119862119906119909 mcs119895(119904) and 119862119906119910 mcs119895(119904) stand for the average values ofcontext

312 Interest Similarity Computation In mobile cloud com-puting environments each user lives in a large network offriends which is called social network It has been reportedthat user interestsrsquo similarity can be leveraged to supportservices and products recommendation [22] A user prefersto choose the items recommended by other users withsimilar interest in a social network Exploring those usersof a high interest similarity with the existing clients couldefficiently enlarge client groups for cloud service providers[22]

Here we use cosine similarity to compute interest similar-ity similar to [22] Interest similarity between users 119906119909 and 119906119910is then defined as the cosine distance between their respectivecloud service invocation sets

sim119868 (119906119909 119906119910) = 100381710038171003817100381710038171003817119868119906119909 cap 1198681199061199101003817100381710038171003817100381710038171003817100381710038171003817100381711986811990611990910038171003817100381710038171003817 sdot 100381710038171003817100381710038171003817119868119906119910100381710038171003817100381710038171003817 (5)

where 119868119906119909 = radic119897119906119909 (119897119906119909 is the number ofmobile cloud serviceswhich have been invoked by 119906119909) and 119868119906119909 cap119868119906119910 is the numberof mobile cloud services which have been coinvoked by 119906119909and 119906119910 If 119897119906119909 = 0 or 119897119906119910 = 0 sim119868(119906119909 119906119910) is undefined

Cloud service invocation is an important factor to deter-mine the interest of users [5] For example user 119906119909 hasinvoked mobile cloud services mcs1 mcs3 mcs6 and mcs7119906119910 has invoked mcs6 and mcs7 and 119906119911 has invoked mcs1 andmcs2 Though 119906119909 119906119910 and 119906119911 do not know each other in realworld and 119906119909 and 119906119910 have a high interest similarity than 119906119909and 119906119911 as they both invoked mcs6 and mcs7 [4]

313 Getting Initial Ranking Results We first predict themissing value 119903119906119909 mcs119895 of mobile cloud service mcs119895 madeby user 119906119898 before making initial ranking results For everymobile cloud service only the Top-119896 similar training useris selected to make missing value prediction Based on thesim119880(119906119909 119906119910) values a set of the Top-k similar training users119873119862mcs119895

(119906119909) over service mcs119895 is identified for target user 119906119909by

119873119862mcs119895(119906119909)

= 119906119910 | 119906119910 isin TS119906119909 sim119880 (119906119909 119906119910) gt 0 119906119909 = 119906119910 (6)

where TS119906119909 is a set of the Top-k similar training users to targetuser 119906119909 and sim119880(119906119909 119906119910) gt 0 excludes the dissimilar userwith negative similar values The value of sim119880(119906119909 119906119910) in (6)is calculated by (2) The missing value 119903119906119909mcs119895 is predicted asfollows

119903119906119909mcs119895= 119903119906119909+ sum119906119910isin119873119862mcs119895

(119906119909)(119903119906119910 mcs119895 minus 119903119906119910) sim119880 (119906119909 119906119910)sum119906119910isin119873119862mcs119895(119906119909)

sim119880 (119906119909 119906119910) (7)

where 119903119906119909 and 119903119906119910 are the average value of service usageexperiences (QoS value or customer rating) of mobile cloudservice made by users 119906119909 and 119906119910 respectively As differentQoS properties have different dimensions and range of valueswe first ensure predicted missing values in the range of [0 1]In [5] QoS properties are classified into two categories ldquocostrdquoand ldquobenefitrdquo For ldquocostrdquo property (response-time) the lowerits value is the greater possibility that a user would chooseit becomes In SSRM all QoS properties are consideredas ldquobenefitrdquo attribute Normalized missing value 119876119906119909 mcs119895 iscomputed as follows

119876119906119909 mcs119895

=

119903119906119909mcs119895 minus min (119903119906119909)max (119903119906119909) minus min (119903119906119909) 119903119906119909mcs119895 isin ldquobenefitrdquo

max (119903119906119909) minus 119903119906119909mcs119895

max (119903119906119909) minus min (119903119906119909) 119903119906119909mcs119895 isin ldquocostrdquo(8)

where min(119903119906119909) and max(119903119906119909) denote the minimum andmaximumQoSproperty value for user119906119909 and they are subjectto the following constraints

min (119903119906119909) = min (119903119906119909mcs119895 | 119895 = 1 119899)max (119903119906119909) = max (119903119906119909 mcs119895 | 119895 = 1 119899) (9)

119876119906119909 mcs119895 is in the range of [0 1] The larger its value isthe more possibility that user 119906119909 would be satisfied with theservice becomes Finally the cloud services are ranked for 119906119909in the order of decreasing 119876119906119909mcs119895 values for cloud serviceselection similar to [4]The basic ranking algorithm is shownin Algorithm 1 The basic ranking algorithm is similar to [3]Compared with [3ndash5] in our ranking algorithm the usersimilarity is measured based on user context informationand user interest In addition the basic ranking algorithm isenhanced in Section 32 In the enhanced ranking algorithmwe calculate the relational degree among cloud services andutilize it to improve the accuracy of ranking results

The basic algorithm ranks the employed mobile cloudservice in 119864 based on the observed historical service usageexperiences The ranking results are stored in 119877119890(119905) which isin the interval [1 |119864|] A smaller value shows a higher rank119905 expresses a mobile cloud service For every cloud servicemcs119895 normalized value 119876119906119909 mcs119895 is calculated Services areranked from high to low by picking up the service 119905 that hasthe maximum 119876119906119909mcs119895 value

Wireless Communications and Mobile Computing 7

Inputan employed service set 119864a full service set 119868an employed similar training user set 119873119862mcs119895

(119906119909)observed service usage experience values 119903119906119910 mcs119895 (119906119910 isin 119873119862mcs119895

(119906119909))Outa service ranking 119865 = 119864while 119865 = Oslash do119905 = argmaxmcs119895isin119865119876119906119909 mcs119895 119877119890(119905) = |119864| minus |119865| + 1119865 = 119865 minus 119905endforeach mcs119895 isin 119868 do

119876119906119909 mcs119895 =

119903119906119909 mcs119895 minus min(119903119906119909 )max(119903119906119909 ) minus min(119903119906119909 ) 119903119906119909 mcs119895 isin ldquobenefitrdquo

max(119903119906119909 ) minus 119903119906119909 mcs119895

max(119903119906119909 ) minus min(119903119906119909 ) 119903119906119909 mcs119895 isin ldquocostrdquo

end119899 = |119868|while 119868 = Oslash do119905 = argmaxmcs119895isin119868119876119906119909 mcs119895(119905) = 119899 minus |119868| + 1119868 = 119868 minus 119905end

end

Algorithm 1 Basic ranking algorithm

32 Improving Accuracy of Ranking Results Through calcu-lating the similar relation among users the missing value119903119906119909mcs119895 is successfully predicted in Section 31 However therelation among cloud services is no doubt an importantevaluation factor of missing value prediction The relatedservices will be probably selected by the same user based onthe intuition [9]Therefore we calculate the relational degreeand utilize it to improve the accuracy of missing value 119903119906119909mcs119895computation in the section The relational degree betweenservice mcs119895 and service mcs119894 is denoted as 119889119895119894 In the paperPropFlow [23] is used to calculate 119889119895119894 PropFlow algorithmcomputes the information flow between services where alarger value indicates tighter relation In order to calculate119889119895119894 we firstly describe the formal definition of service relationgraph

Definition 2 (service relation graph) Given a set of servicesMCS and a totally ordered domain of weights 119882 a servicerelation graph is a weighted undirected graph 119866 = (MCS 119864)The edge (mcs119894mcs119895 119908119894119895) in the set 119864 isin 119880 times 119880 times 119882 encodesthe link weight119908119894119895 (119908119894119895 ge 0) between service mcs119894 and servicemcs119895 An edge depicts the direct link relationship betweenmcs119894 and mcs119895 As 119866 is undirected 119908119894119895 = 119908119895119894

Equation (10) shows how to calculate 119889119895119894119889119895119894 = Input119895 lowast 119908119895119894sum119896isin119873(119895) 119908119895119896 (10)

Ｇ＝Ｍℎ

Ｇ＝ＭiＧ＝Ｍl

Ｇ＝Ｍe

Ｇ＝Ｍf

Ｇ＝Ｍj Ｇ＝Ｍk

wkl = 1 wli = 5

wle = 1

wie = 1wke = 1

wkℎ = 2

wjk = 3

wjf = 1

Figure 3 An example of service relation graph

where initial input is regarded as 1 and 119873(119895) is the set ofneighbors of mcs119895

Figure 3 shows an example of relation degree computa-tion where there are six kinds of mobile cloud services mcs119894and mcs119895 are indirectly linked We assume mcs119895 is startingnode

Four paths can reach mcs119894 from mcs119895 but only twopaths are considered for computation based on PropFlow

8 Wireless Communications and Mobile Computing

algorithm They are mcs119895 rarr mcs119896 rarr mcs119897 rarr mcs119894 andmcs119895 rarr mcs119896 rarr mcs119890 rarr mcs119894

First we compute 119889119895119896 The sum of link weights of mcs119895and its neighbor is 4 119889119895119896 is computed as

119889119895119896 = 1 lowast 3(1 + 3) = 1 lowast 34 = 34 (11)

119889119896119897 is computed as

119889119896119897 = 119889119895119896 lowast 1(1 + 1 + 2) = 34 lowast 14 = 316 (12)

With the same method 119889119895119894 is computed as

119889119895119894 = 119889119897119894 + 119889119890119894 = 119889119896119897 lowast 55 + 119889119896119890 lowast 11 = 38 (13)

More details of calculating 119889119895119894 are shown in [24]Next we assume service mcs119894 is invoked recently by user119906119909 119889119895119894 is incorporated into (7) and then the missing value119903119906119909mcs119895 is computed with (14) in basic ranking algorithm

Finally the ultimate ranking results are got

119903119906119909 mcs119895 = (119903119906119909+ sum119906119910isin119873119862mcs119895

(119906119909)(119903119906119910 mcs119895 minus 119903119906119910) sim119880 (119906119909 119906119910)sum119906119910isin119873119862mcs119895(119906119909)

sim119880 (119906119909 119906119910) )sdot (1 + 119889119895119894)

(14)

In SSRM we only select trusted mobile cloud serviceas candidates Therefore a set of trusted candidates areidentified for 119906119909 by

MCS119909 = mcs119895 | trustmcs119895 gt 120583119909 119895 = 1 119873 (15)

where trustmcs119895 denotes the trustworthiness of mobile cloudservice mcs119895 and 120583119909 expresses the trust threshold decidedby user 119906119909 Many existing methods [4ndash6] can be applied tocompute trustmcs119895 Here how to compute trustmcs119895 is not tobe discussed We omit the details for brevity

4 Experimental Study

In this section in order to evaluate the effectiveness ofSSRM a series of test scenarios are developed To studythe prediction performance we compare SSRM with othertwo traditional approaches item-based CF approach (IBCF)and user-based CF approach (UBCF) SSRM IBCF andUBCF are rating-oriented methods which rank the cloudservices based on the predicted QoS values Since there isno suitable real data supporting mobile cloud computingsimulation we use ns3 simulator to generate the experimentaldataset Firstly we generated cloud service invocation codes

by Axis2 [25] a Java-based open source package for cloudservices Then we simulate mobile cloud service on ns3[26] based on the invocation codes On ns3 the creationof servers mobile users and service model was easier thanwith other network simulators In the simulating process ofdataset the real-world web service QoS dataset from WS-DREAM team [3] is referenced by usThe detailed real-worldQoS values are publicly released online [27] which makesour experimental evaluations reproducible We simulate 100mobile users 25 servers and 300 services The value of linkweight between two services is randomly generated which isin the range of [1 5] In order to conduct our experimentsrealistically the changes of user context information aresimulated Three types of context are considered includingbandwidth memory and battery capacity The bandwidthvalue of a mobile user is in the range of 4Mndash8MThe valuesof memory and battery capacity will decline over timeThesevalues are in the range of 100000000ndash2100000000MB and05ndash2KVA

We normalize the context information by

119862nor119906119909mcs119895 (V) = 119886 tan (119862119906119909 mcs119895 (V)) lowast 2120587 (16)

where 119862nor119906119909mcs119895(V) denotes the normalized context value of

user 119906119909 over cloud service mcs119895 Inverse tangent function isapplied to normalize the raw data 119862119906119909 mcs119895(V) Then thesenormalized property values are used to calculate contextsimilarity with (4) Once the context simulation is finishedthe target user 119906119909 will obtain context similarity results foreach cloud service

Similar to [3] QoS properties include throughput andresponse-time that are in the range of 0ndash1000 kbps and 0ndash20second respectively Most of the throughput and response-time values fall between 5ndash40 kbps and 01ndash08 seconds Inorder to simply the experimental process 200 service QoSrecords are randomly selected and two 100 times 200 user-servicematrices are constructed The two matrices include through-put and response-time respectivelyUser-servicematrices areseparated into two parts training set (80 historical usageexperiences in the matrix) and test set (the remaining 20usage experiences) Each entry in the matrix is the QoSvalue (eg response-time or throughput) of a mobile cloudservice observed by a user The experiments employ the QoSvalues of response-time and throughput to rank the servicesindependently Table 1 shows descriptions of the obtainedpractical cloud service QoS values and context information

We use Mean Absolute Error (MAE) and Root MeanSquare Error (RMSE) as the metric to evaluate predictionperformance of the proposed approach in comparison withother approaches MAE and RMSE are defined as

MAE = sum119906119909 mcs119895

100381610038161003816100381610038161003816119903act119906119909mcs119895 minus 119903pre119906119909mcs119895100381610038161003816100381610038161003816119897pre

RMSE = radic sum119906119909 mcs119895 (119903act119906119909mcs119895 minus 119903pre119906119909 mcs119895)2119897pre (17)

Wireless Communications and Mobile Computing 9

Response-time

02

025

03

035

04

045

05

055

06

065

07

MA

E

01 02 03 04 05 06 07 08 09 100Weighting factor

(a) MAE values

Response-time

03

035

04

045

05

055

06

065

07

075

08

RMSE

01 02 03 04 05 06 07 08 09 100Weighting factor

(b) RMSE values

Figure 4 Impact of weighting factor 120582

Table 1 Mobile cloud service QoS dataset and context informationdescriptions

Statistics ValueNumber of mobile cloud serviceinvocations 180000

Number of service users 100Number of mobile cloud services 200Minimum response-time value 0004 sMaximum response-time value 20 sMean of response-time 110 sStandard deviation ofresponse-time 226 s

Minimum throughput value 01 kbpsMaximum throughput value 1000 kbpsMean of throughput 2546 kbpsStandard deviation of throughput 4803 kbpsBandwidth 4Mndash8MMemory 100000000MBndash2100000000MBBattery capacity 05 KVAndash2KVA

where 119903act119906119909mcs119895 and 119903pre119906119909mcs119895 denote the actual QoS valueand predicted value respectively 119897pre is the number ofpredicted QoS values Smaller values of MAE and RMSEindicate better results

41 Impact of 120582 120582 is weighting factor in (2) 120582 determineshow much similarity measure relies on context and interestIn the experiment we vary the weighting factor 120582 from 0 to 1in increment of 01 The Top-k is set to 5 Matrix density is setto 10 Matrix density means the percentage of selected QoSentries that are used to predict the missing QoS value Theexperimental results are shown in Figure 4 As the value of 120582increases MAE firstly decreases and then quickly increasesWhen 120582 = 04 our results have the best value As shown inFigure 4(b) RMSE presents similar trend

42 Performance Comparison of SSRM IBCF and UBCFTo compare the performance of SSRM IBCF and UBCFsimilarity computation methods we implement IBCF andUBCF methods

Item-Based CF Approach (IBCF) We apply PCC to calculatesimilarities between users and predicted QoS values based onsimilar users The user similarity is calculated by

sim (119906119909 119906119910) = summcs119895isinmcs119906119909119906119910(119903119906119909 mcs119895 minus 119903119906119909) (119903119906119910 mcs119895 minus 119903119906119910)

radicsummcs119895isinmcs119906119909119906119910(119903119906119909mcs119895 minus 119903119906119909)2radicsummcs119895isinmcs119906119909119906119910

(119903119906119910mcs119895 minus 119903119906119910)2 (18)

10 Wireless Communications and Mobile Computing

where mcs119906119909119906119910 is the set of mobile cloud services that havebeen coinvoked by 119906119909 and 119906119910 119903119906119909and 119903119906119910 are average QoSvalues of cloud service invoked by 119906119909 and 119906119910 respectively

User-Based CF Approach (UBCF) We apply PCC to calculatesimilarities between services and predicted QoS values basedon similar services The service similarity is calculated by

sim (mcs119895mcs119894) = sum119906119909isin119906mcs119895mcs119894(119903119906119909mcs119895 minus 119903mcs119895) (119903119906119909 mcs119894 minus 119903mcs119894)

radicsum119906119909isin119906mcs119895mcs119894(119903119906119909mcs119895 minus 119903mcs119895)2radicsum119906119909isin119906mcs119895mcs119894

(119903119906119909mcs119894 minus 119903mcs119894)2 (19)

where 119906mcs119895mcs119894 is the set of users who have invoked the cloudservices mcs119895 andmcs119894 119903mcs119895 and 119903mcs119894 are average QoS valuesof cloud services mcs119895 and mcs119894 made by 119906119909

In the experiments trustworthiness of cloud service hasno influence on similarity computation and service decisionin the experiments In order to simplify the experimentalprocess we consider all cloud services are trustworthyWeighting factor 120582 is set to 04 We firstly study the effectof neighbor size 119896 Matrix density is set to 10 We change119896 from 5 to 30 in increment of 5 Figure 5 shows theexperimental results for response-time and throughput

As shown in Figure 5 the prediction performance ofSSRM outperforms other two approaches The performanceof IBCF is similar to UBCF As the values of 119896 increase betteraccuracy can be achieved This indicates that more similarneighbor records can provide more information for missingvalue prediction However when 119896 is larger than 25 MAEand RSME fail to drop with the value of 119896 increasing Themain reason is the limited number of similar neighbors Theobservations also show that RMSE has the similar trend butwith larger fluctuations

Next we investigate the impact of matrix density Thematrix density varies from 10 to 50 in increment of10 Similar user size k is set to 5 in this experimentFigure 6 shows the experimental results for response-timeand throughput

As presented in Figure 6 the prediction performance isalso enhanced with the value of matrix density increasingThe main reason is that denser user-service matrix providesmore information for missing value prediction NARM hasthe better performance than LBCF and UBCF under allexperimental setting consistently since NARM considerscontext similarity and the relational degree among cloudservices The experimental results of LBCF and UBCF aresimilar in this experiment since these two similarity compu-tation methods are similar to each other and are both rating-oriented

43 Performance Comparison with Other Popular ApproachesTo study the prediction performance of SSRM we com-pare SSRM with three existing QoS properties predictionapproaches CloudRank2 [3] TECSS [4] and JV-PCC [5]

The size of Top-119896 similar service is an important factorin CF approach which determines how many neighborsrsquohistorical records are employed to generate predictions [5]Therefore the impact of matrix density is not considered in

the experiments We set the density to 10 We vary 119896 from 5to 30 in increment of 5 In order to simplify the experimentalprocess we consider all cloud services are trustworthyFigure 7 shows the experimental results for response-timeand throughput Under the same simulation condition SSRMsignificantly outperforms CloudRank2 TECSS and JV-PCCThe observations also suggest that better accuracy can beachieved by our model when more historical records areavailable in the service selection studyThemain reason is thatcomputing context similarity can improve the performance ofprediction evaluation in MCC environments

Figures 7(a) and 7(b) depict the MAE fractions of SSRMCloudRank2 TECSS and JV-PCC for response-time andthroughput while Figures 7(c) and 7(d) depict the RMSEfractions It can be observed that SSRM achieves smallerMAE and RMSE consistently than other approaches for bothresponse-time and throughput

5 Conclusions and Future

In the paper we propose a novel service selection and recom-mendation model (SSRM) for mobile cloud computing envi-ronmentsWhen calculating user similarity context informa-tion and interest are considered Through utilizing relationaldegree to improve the accuracy of ranking result our serviceselection and recommendation method succeed in obtainingthe ultimate service ranking results In addition SSRM onlyselects trusted mobile cloud service as candidates Thereforea set of trusted candidates are identified for target user Theexperimental results show that our approach significantlyimproves the prediction performance as comparedwith othertwo traditional approaches item-based CF approach (IBCF)and user-based CF approach (UBCF)

There are some disadvantages in our SSRM Firstly weexploit node distance to compute context similarity Thereis an underlying assumption in this exploitation each ofthe two adjacent nodes has equal semantic distance orgranularity of nodes in each level is identicalThis underlyingassumption is not true in some scenarios and it deterioratesthe performance of similarity evaluation Secondly the trustthreshold is used to select trusted mobile cloud service ascandidates However it cannot detect and exclude maliciousQoS values provided by users Thirdly in our experimentsthere are only three types of context information that areconsidered

Therefore in the future we will study the methods toimprove context similarity measurement We will study how

Wireless Communications and Mobile Computing 11

Response-time

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E10 25 3015 205

k

(b)Response-time

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

(c)

Throughput

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

(d)

Figure 5 Impact of neighbor size 119896 comparison with SSRM IBCF and UBCF

12 Wireless Communications and Mobile Computing

Response-time

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E

20 50 6030 4010Matrix density ()

(a)

Throughput

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E20 30 40 50 6010

Matrix density ()

(b)

Response-time

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 50 6030 4010Matrix density ()

(c)

Throughput

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 30 40 50 6010Matrix density ()

(d)

Figure 6 Impact of matrix density comparison with SSRM IBCF and UBCF

Wireless Communications and Mobile Computing 13

Response-time

JV-PCCTECSSSSRM

CloudRank2

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

02

025

03

035

04

045

05

055

06

065

07

MA

E10 15 20 25 305

k

JV-PCCTECSSSSRM

CloudRank2

(b)

Response-time

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

JV-PCCTECSSSSRM

CloudRank2

(c)

Throughput

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

JV-PCCTECSSSSRM

CloudRank2

(d)

Figure 7 Impact of neighbor size 119896 comparison with other popular approaches

14 Wireless Communications and Mobile Computing

to select themost trustworthy cloud service of certain type forthe active users based on multidimensional trust evidenceWe also will conduct more experimental investigations thatdeal with the impact of different context changes on serviceselection Moreover we will improve the ranking accuracyof our approaches by exploiting additional techniques (com-bining rating-oriented approaches and ranking-orientedapproaches matrix factorization intelligent optimizationalgorithms etc)

Conflicts of Interest

The author declares that they have no conflicts of interest

Acknowledgments

The work in this paper has been supported by NationalNatural Science Foundation of China (Program no 71501156)and China Postdoctoral Science Foundation (Program no2014M560796) and Shanxi Provincial EducationDepartment(Program no 15JK1679)

References

[1] White PaperMobile CloudComputing Solution Brief AEPONA2010

[2] httpsenwikipediaorgwikiCloudlet[3] Z Zheng X Wu Y Zhang M R Lyu and J Wang ldquoQoS

ranking prediction for cloud servicesrdquo IEEE Transactions onParallel and Distributed Systems vol 24 no 6 pp 1213ndash12222013

[4] Y Pan SDingW Fan J Li and S Yang ldquoTrust-enhanced cloudservice selection model based on QoS analysisrdquo PLoS ONE vol10 no 11 Article ID e0143448 2015

[5] S Ding C-Y Xia K-L Zhou S-L Yang and J S Shang ldquoDeci-sion support for personalized cloud service selection throughmulti-attribute trustworthiness evaluationrdquo PLoS ONE vol 9no 6 Article ID e107821 2014

[6] S K Garg S Versteeg and R Buyya ldquoA framework for rankingof cloud computing servicesrdquo Future Generation ComputerSystems vol 29 no 4 pp 1012ndash1023 2013

[7] H Ma Z Hu K Li and H Zhang ldquoToward trustworthy cloudservice selection a time-aware approach using interval neutro-sophic setrdquo Journal of Parallel and Distributed Computing vol96 pp 75ndash94 2016

[8] A Ahmed and A S Sabyasachi ldquoMobile cloud service selectionusing back propagation neural networkrdquo International Journalof Computer Applications vol 129 no 14 pp 1ndash5 2015

[9] L Guo J Ma Z-M Chen and H-R Jiang ldquoIncorporatingitem relations for social recommendationrdquo Chinese Journal ofComputers vol 37 no 1 pp 219ndash228 2014

[10] Z Yang B Wu K Zheng X Wang and L Lei ldquoA survey ofcollaborative filtering-based recommender systems for mobileinternet applicationsrdquo IEEE Access vol 4 pp 3273ndash3287 2016

[11] Markets and Markets httpwwwmarketsandmarketscomPressReleasesmobile-cloudasp

[12] M Whaiduzzaman A Gani N B Anuar M Shiraz MN Haque and I T Haque ldquoCloud service selection usingmulticriteria decision analysisrdquoTheScientificWorld Journal vol2014 Article ID 459375 10 pages 2014

[13] K Tserpes F Aisopos D Kyriazis and T Varvarigou ldquoArecommender mechanism for service selection in service-oriented environmentsrdquo Future Generation Computer Systemsvol 28 no 8 pp 1285ndash1294 2012

[14] Z U Rehman O K Hussain and F K Hussain ldquoParallelcloud service selection and ranking based on QoS historyrdquoInternational Journal of Parallel Programming vol 42 no 5 pp820ndash852 2014

[15] W-J Fan S-L YangH Perros and J Pei ldquoAmulti-dimensionaltrust-aware cloud service selection mechanism based on evi-dential reasoning approachrdquo International Journal of Automa-tion and Computing vol 12 no 2 pp 208ndash219 2015

[16] X G Wang J Cao and Y Xiang ldquoDynamic cloud serviceselection using an adaptive learning mechanism in multi-cloudcomputingrdquo The Journal of Systems and Software vol 100 pp195ndash210 2015

[17] C T Do N H Tran E-N Huh C S Hong D Niyato andZ Han ldquoDynamics of service selection and provider pricinggame in heterogeneous cloud marketrdquo Journal of Network andComputer Applications vol 69 pp 152ndash165 2015

[18] R Pal and P Hui ldquoEconomic models for cloud service marketspricing and capacity planningrdquo Theoretical Computer Sciencevol 496 pp 113ndash124 2013

[19] A Ezenwoke O Daramola and M Adigun ldquoTowards avisualization framework for service selection in cloud e-marketplacesrdquo in Proceedings of the 2017 IEEE InternationalConference on Web Services (ICWS rsquo17) pp 828ndash835 HonoluluHI USA June 2017

[20] M Tang X Dai J Liu and J Chen ldquoTowards a trust evaluationmiddleware for cloud service selectionrdquo Future GenerationComputer Systems vol 74 pp 302ndash312 2017

[21] A K Dey ldquoUnderstanding and using contextrdquo Personal andUbiquitous Computing vol 5 no 1 pp 4ndash7 2001

[22] X Han L Wang S Park A Cuevas and N Crespi ldquoAlikepeople alike interests A large-scale study on interest similarityin social networksrdquo in Proceedings of the 2014 IEEEACM Inter-national Conference on Advances in Social Networks Analysisand Mining (ASONAM rsquo14) pp 491ndash496 August 2014

[23] R N Lichtenwalter J T Lussier and N V Chawla ldquoNewperspectives and methods in link predictionrdquo in Proceedings ofthe 16th ACM SIGKDD International Conference on KnowledgeDiscovery and Data Mining (KDD rsquo10) pp 243ndash252 July 2010

[24] L Munasinghe and R Ichise ldquoLink prediction in social net-works using information flow via active linksrdquo IEICE Transac-tion on Information and Systems vol E96-D no 7 pp 1495ndash1502 2013

[25] Axis httpaxisapacheorgaxis2[26] ns3 httpswwwnsnamorgns3[27] Dataset httpwwwzibinzhengcomtpds

6 Wireless Communications and Mobile Computing

where sim119862mcs119895(119906119909 119906119910) expresses the context similarity

between 119906119909 and 119906119910 and it is in the interval of [minus1 1] and119862119906119909 mcs119895(119904) and 119862119906119910 mcs119895(119904) stand for the average values ofcontext

312 Interest Similarity Computation In mobile cloud com-puting environments each user lives in a large network offriends which is called social network It has been reportedthat user interestsrsquo similarity can be leveraged to supportservices and products recommendation [22] A user prefersto choose the items recommended by other users withsimilar interest in a social network Exploring those usersof a high interest similarity with the existing clients couldefficiently enlarge client groups for cloud service providers[22]

Here we use cosine similarity to compute interest similar-ity similar to [22] Interest similarity between users 119906119909 and 119906119910is then defined as the cosine distance between their respectivecloud service invocation sets

sim119868 (119906119909 119906119910) = 100381710038171003817100381710038171003817119868119906119909 cap 1198681199061199101003817100381710038171003817100381710038171003817100381710038171003817100381711986811990611990910038171003817100381710038171003817 sdot 100381710038171003817100381710038171003817119868119906119910100381710038171003817100381710038171003817 (5)

where 119868119906119909 = radic119897119906119909 (119897119906119909 is the number ofmobile cloud serviceswhich have been invoked by 119906119909) and 119868119906119909 cap119868119906119910 is the numberof mobile cloud services which have been coinvoked by 119906119909and 119906119910 If 119897119906119909 = 0 or 119897119906119910 = 0 sim119868(119906119909 119906119910) is undefined

Cloud service invocation is an important factor to deter-mine the interest of users [5] For example user 119906119909 hasinvoked mobile cloud services mcs1 mcs3 mcs6 and mcs7119906119910 has invoked mcs6 and mcs7 and 119906119911 has invoked mcs1 andmcs2 Though 119906119909 119906119910 and 119906119911 do not know each other in realworld and 119906119909 and 119906119910 have a high interest similarity than 119906119909and 119906119911 as they both invoked mcs6 and mcs7 [4]

313 Getting Initial Ranking Results We first predict themissing value 119903119906119909 mcs119895 of mobile cloud service mcs119895 madeby user 119906119898 before making initial ranking results For everymobile cloud service only the Top-119896 similar training useris selected to make missing value prediction Based on thesim119880(119906119909 119906119910) values a set of the Top-k similar training users119873119862mcs119895

(119906119909) over service mcs119895 is identified for target user 119906119909by

119873119862mcs119895(119906119909)

= 119906119910 | 119906119910 isin TS119906119909 sim119880 (119906119909 119906119910) gt 0 119906119909 = 119906119910 (6)

where TS119906119909 is a set of the Top-k similar training users to targetuser 119906119909 and sim119880(119906119909 119906119910) gt 0 excludes the dissimilar userwith negative similar values The value of sim119880(119906119909 119906119910) in (6)is calculated by (2) The missing value 119903119906119909mcs119895 is predicted asfollows

119903119906119909mcs119895= 119903119906119909+ sum119906119910isin119873119862mcs119895

(119906119909)(119903119906119910 mcs119895 minus 119903119906119910) sim119880 (119906119909 119906119910)sum119906119910isin119873119862mcs119895(119906119909)

sim119880 (119906119909 119906119910) (7)

where 119903119906119909 and 119903119906119910 are the average value of service usageexperiences (QoS value or customer rating) of mobile cloudservice made by users 119906119909 and 119906119910 respectively As differentQoS properties have different dimensions and range of valueswe first ensure predicted missing values in the range of [0 1]In [5] QoS properties are classified into two categories ldquocostrdquoand ldquobenefitrdquo For ldquocostrdquo property (response-time) the lowerits value is the greater possibility that a user would chooseit becomes In SSRM all QoS properties are consideredas ldquobenefitrdquo attribute Normalized missing value 119876119906119909 mcs119895 iscomputed as follows

119876119906119909 mcs119895

=

119903119906119909mcs119895 minus min (119903119906119909)max (119903119906119909) minus min (119903119906119909) 119903119906119909mcs119895 isin ldquobenefitrdquo

max (119903119906119909) minus 119903119906119909mcs119895

max (119903119906119909) minus min (119903119906119909) 119903119906119909mcs119895 isin ldquocostrdquo(8)

where min(119903119906119909) and max(119903119906119909) denote the minimum andmaximumQoSproperty value for user119906119909 and they are subjectto the following constraints

min (119903119906119909) = min (119903119906119909mcs119895 | 119895 = 1 119899)max (119903119906119909) = max (119903119906119909 mcs119895 | 119895 = 1 119899) (9)

119876119906119909 mcs119895 is in the range of [0 1] The larger its value isthe more possibility that user 119906119909 would be satisfied with theservice becomes Finally the cloud services are ranked for 119906119909in the order of decreasing 119876119906119909mcs119895 values for cloud serviceselection similar to [4]The basic ranking algorithm is shownin Algorithm 1 The basic ranking algorithm is similar to [3]Compared with [3ndash5] in our ranking algorithm the usersimilarity is measured based on user context informationand user interest In addition the basic ranking algorithm isenhanced in Section 32 In the enhanced ranking algorithmwe calculate the relational degree among cloud services andutilize it to improve the accuracy of ranking results

The basic algorithm ranks the employed mobile cloudservice in 119864 based on the observed historical service usageexperiences The ranking results are stored in 119877119890(119905) which isin the interval [1 |119864|] A smaller value shows a higher rank119905 expresses a mobile cloud service For every cloud servicemcs119895 normalized value 119876119906119909 mcs119895 is calculated Services areranked from high to low by picking up the service 119905 that hasthe maximum 119876119906119909mcs119895 value

Wireless Communications and Mobile Computing 7

Inputan employed service set 119864a full service set 119868an employed similar training user set 119873119862mcs119895

(119906119909)observed service usage experience values 119903119906119910 mcs119895 (119906119910 isin 119873119862mcs119895

(119906119909))Outa service ranking 119865 = 119864while 119865 = Oslash do119905 = argmaxmcs119895isin119865119876119906119909 mcs119895 119877119890(119905) = |119864| minus |119865| + 1119865 = 119865 minus 119905endforeach mcs119895 isin 119868 do

119876119906119909 mcs119895 =

119903119906119909 mcs119895 minus min(119903119906119909 )max(119903119906119909 ) minus min(119903119906119909 ) 119903119906119909 mcs119895 isin ldquobenefitrdquo

max(119903119906119909 ) minus 119903119906119909 mcs119895

max(119903119906119909 ) minus min(119903119906119909 ) 119903119906119909 mcs119895 isin ldquocostrdquo

end119899 = |119868|while 119868 = Oslash do119905 = argmaxmcs119895isin119868119876119906119909 mcs119895(119905) = 119899 minus |119868| + 1119868 = 119868 minus 119905end

end

Algorithm 1 Basic ranking algorithm

32 Improving Accuracy of Ranking Results Through calcu-lating the similar relation among users the missing value119903119906119909mcs119895 is successfully predicted in Section 31 However therelation among cloud services is no doubt an importantevaluation factor of missing value prediction The relatedservices will be probably selected by the same user based onthe intuition [9]Therefore we calculate the relational degreeand utilize it to improve the accuracy of missing value 119903119906119909mcs119895computation in the section The relational degree betweenservice mcs119895 and service mcs119894 is denoted as 119889119895119894 In the paperPropFlow [23] is used to calculate 119889119895119894 PropFlow algorithmcomputes the information flow between services where alarger value indicates tighter relation In order to calculate119889119895119894 we firstly describe the formal definition of service relationgraph

Definition 2 (service relation graph) Given a set of servicesMCS and a totally ordered domain of weights 119882 a servicerelation graph is a weighted undirected graph 119866 = (MCS 119864)The edge (mcs119894mcs119895 119908119894119895) in the set 119864 isin 119880 times 119880 times 119882 encodesthe link weight119908119894119895 (119908119894119895 ge 0) between service mcs119894 and servicemcs119895 An edge depicts the direct link relationship betweenmcs119894 and mcs119895 As 119866 is undirected 119908119894119895 = 119908119895119894

Equation (10) shows how to calculate 119889119895119894119889119895119894 = Input119895 lowast 119908119895119894sum119896isin119873(119895) 119908119895119896 (10)

Ｇ＝Ｍℎ

Ｇ＝ＭiＧ＝Ｍl

Ｇ＝Ｍe

Ｇ＝Ｍf

Ｇ＝Ｍj Ｇ＝Ｍk

wkl = 1 wli = 5

wle = 1

wie = 1wke = 1

wkℎ = 2

wjk = 3

wjf = 1

Figure 3 An example of service relation graph

where initial input is regarded as 1 and 119873(119895) is the set ofneighbors of mcs119895

Figure 3 shows an example of relation degree computa-tion where there are six kinds of mobile cloud services mcs119894and mcs119895 are indirectly linked We assume mcs119895 is startingnode

Four paths can reach mcs119894 from mcs119895 but only twopaths are considered for computation based on PropFlow

8 Wireless Communications and Mobile Computing

algorithm They are mcs119895 rarr mcs119896 rarr mcs119897 rarr mcs119894 andmcs119895 rarr mcs119896 rarr mcs119890 rarr mcs119894

First we compute 119889119895119896 The sum of link weights of mcs119895and its neighbor is 4 119889119895119896 is computed as

119889119895119896 = 1 lowast 3(1 + 3) = 1 lowast 34 = 34 (11)

119889119896119897 is computed as

119889119896119897 = 119889119895119896 lowast 1(1 + 1 + 2) = 34 lowast 14 = 316 (12)

With the same method 119889119895119894 is computed as

119889119895119894 = 119889119897119894 + 119889119890119894 = 119889119896119897 lowast 55 + 119889119896119890 lowast 11 = 38 (13)

More details of calculating 119889119895119894 are shown in [24]Next we assume service mcs119894 is invoked recently by user119906119909 119889119895119894 is incorporated into (7) and then the missing value119903119906119909mcs119895 is computed with (14) in basic ranking algorithm

Finally the ultimate ranking results are got

119903119906119909 mcs119895 = (119903119906119909+ sum119906119910isin119873119862mcs119895

(119906119909)(119903119906119910 mcs119895 minus 119903119906119910) sim119880 (119906119909 119906119910)sum119906119910isin119873119862mcs119895(119906119909)

sim119880 (119906119909 119906119910) )sdot (1 + 119889119895119894)

(14)

In SSRM we only select trusted mobile cloud serviceas candidates Therefore a set of trusted candidates areidentified for 119906119909 by

MCS119909 = mcs119895 | trustmcs119895 gt 120583119909 119895 = 1 119873 (15)

where trustmcs119895 denotes the trustworthiness of mobile cloudservice mcs119895 and 120583119909 expresses the trust threshold decidedby user 119906119909 Many existing methods [4ndash6] can be applied tocompute trustmcs119895 Here how to compute trustmcs119895 is not tobe discussed We omit the details for brevity

4 Experimental Study

In this section in order to evaluate the effectiveness ofSSRM a series of test scenarios are developed To studythe prediction performance we compare SSRM with othertwo traditional approaches item-based CF approach (IBCF)and user-based CF approach (UBCF) SSRM IBCF andUBCF are rating-oriented methods which rank the cloudservices based on the predicted QoS values Since there isno suitable real data supporting mobile cloud computingsimulation we use ns3 simulator to generate the experimentaldataset Firstly we generated cloud service invocation codes

by Axis2 [25] a Java-based open source package for cloudservices Then we simulate mobile cloud service on ns3[26] based on the invocation codes On ns3 the creationof servers mobile users and service model was easier thanwith other network simulators In the simulating process ofdataset the real-world web service QoS dataset from WS-DREAM team [3] is referenced by usThe detailed real-worldQoS values are publicly released online [27] which makesour experimental evaluations reproducible We simulate 100mobile users 25 servers and 300 services The value of linkweight between two services is randomly generated which isin the range of [1 5] In order to conduct our experimentsrealistically the changes of user context information aresimulated Three types of context are considered includingbandwidth memory and battery capacity The bandwidthvalue of a mobile user is in the range of 4Mndash8MThe valuesof memory and battery capacity will decline over timeThesevalues are in the range of 100000000ndash2100000000MB and05ndash2KVA

We normalize the context information by

119862nor119906119909mcs119895 (V) = 119886 tan (119862119906119909 mcs119895 (V)) lowast 2120587 (16)

where 119862nor119906119909mcs119895(V) denotes the normalized context value of

user 119906119909 over cloud service mcs119895 Inverse tangent function isapplied to normalize the raw data 119862119906119909 mcs119895(V) Then thesenormalized property values are used to calculate contextsimilarity with (4) Once the context simulation is finishedthe target user 119906119909 will obtain context similarity results foreach cloud service

Similar to [3] QoS properties include throughput andresponse-time that are in the range of 0ndash1000 kbps and 0ndash20second respectively Most of the throughput and response-time values fall between 5ndash40 kbps and 01ndash08 seconds Inorder to simply the experimental process 200 service QoSrecords are randomly selected and two 100 times 200 user-servicematrices are constructed The two matrices include through-put and response-time respectivelyUser-servicematrices areseparated into two parts training set (80 historical usageexperiences in the matrix) and test set (the remaining 20usage experiences) Each entry in the matrix is the QoSvalue (eg response-time or throughput) of a mobile cloudservice observed by a user The experiments employ the QoSvalues of response-time and throughput to rank the servicesindependently Table 1 shows descriptions of the obtainedpractical cloud service QoS values and context information

We use Mean Absolute Error (MAE) and Root MeanSquare Error (RMSE) as the metric to evaluate predictionperformance of the proposed approach in comparison withother approaches MAE and RMSE are defined as

MAE = sum119906119909 mcs119895

100381610038161003816100381610038161003816119903act119906119909mcs119895 minus 119903pre119906119909mcs119895100381610038161003816100381610038161003816119897pre

RMSE = radic sum119906119909 mcs119895 (119903act119906119909mcs119895 minus 119903pre119906119909 mcs119895)2119897pre (17)

Wireless Communications and Mobile Computing 9

Response-time

02

025

03

035

04

045

05

055

06

065

07

MA

E

01 02 03 04 05 06 07 08 09 100Weighting factor

(a) MAE values

Response-time

03

035

04

045

05

055

06

065

07

075

08

RMSE

01 02 03 04 05 06 07 08 09 100Weighting factor

(b) RMSE values

Figure 4 Impact of weighting factor 120582

Table 1 Mobile cloud service QoS dataset and context informationdescriptions

Statistics ValueNumber of mobile cloud serviceinvocations 180000

Number of service users 100Number of mobile cloud services 200Minimum response-time value 0004 sMaximum response-time value 20 sMean of response-time 110 sStandard deviation ofresponse-time 226 s

Minimum throughput value 01 kbpsMaximum throughput value 1000 kbpsMean of throughput 2546 kbpsStandard deviation of throughput 4803 kbpsBandwidth 4Mndash8MMemory 100000000MBndash2100000000MBBattery capacity 05 KVAndash2KVA

where 119903act119906119909mcs119895 and 119903pre119906119909mcs119895 denote the actual QoS valueand predicted value respectively 119897pre is the number ofpredicted QoS values Smaller values of MAE and RMSEindicate better results

41 Impact of 120582 120582 is weighting factor in (2) 120582 determineshow much similarity measure relies on context and interestIn the experiment we vary the weighting factor 120582 from 0 to 1in increment of 01 The Top-k is set to 5 Matrix density is setto 10 Matrix density means the percentage of selected QoSentries that are used to predict the missing QoS value Theexperimental results are shown in Figure 4 As the value of 120582increases MAE firstly decreases and then quickly increasesWhen 120582 = 04 our results have the best value As shown inFigure 4(b) RMSE presents similar trend

42 Performance Comparison of SSRM IBCF and UBCFTo compare the performance of SSRM IBCF and UBCFsimilarity computation methods we implement IBCF andUBCF methods

Item-Based CF Approach (IBCF) We apply PCC to calculatesimilarities between users and predicted QoS values based onsimilar users The user similarity is calculated by

sim (119906119909 119906119910) = summcs119895isinmcs119906119909119906119910(119903119906119909 mcs119895 minus 119903119906119909) (119903119906119910 mcs119895 minus 119903119906119910)

radicsummcs119895isinmcs119906119909119906119910(119903119906119909mcs119895 minus 119903119906119909)2radicsummcs119895isinmcs119906119909119906119910

(119903119906119910mcs119895 minus 119903119906119910)2 (18)

10 Wireless Communications and Mobile Computing

where mcs119906119909119906119910 is the set of mobile cloud services that havebeen coinvoked by 119906119909 and 119906119910 119903119906119909and 119903119906119910 are average QoSvalues of cloud service invoked by 119906119909 and 119906119910 respectively

User-Based CF Approach (UBCF) We apply PCC to calculatesimilarities between services and predicted QoS values basedon similar services The service similarity is calculated by

sim (mcs119895mcs119894) = sum119906119909isin119906mcs119895mcs119894(119903119906119909mcs119895 minus 119903mcs119895) (119903119906119909 mcs119894 minus 119903mcs119894)

radicsum119906119909isin119906mcs119895mcs119894(119903119906119909mcs119895 minus 119903mcs119895)2radicsum119906119909isin119906mcs119895mcs119894

(119903119906119909mcs119894 minus 119903mcs119894)2 (19)

where 119906mcs119895mcs119894 is the set of users who have invoked the cloudservices mcs119895 andmcs119894 119903mcs119895 and 119903mcs119894 are average QoS valuesof cloud services mcs119895 and mcs119894 made by 119906119909

In the experiments trustworthiness of cloud service hasno influence on similarity computation and service decisionin the experiments In order to simplify the experimentalprocess we consider all cloud services are trustworthyWeighting factor 120582 is set to 04 We firstly study the effectof neighbor size 119896 Matrix density is set to 10 We change119896 from 5 to 30 in increment of 5 Figure 5 shows theexperimental results for response-time and throughput

As shown in Figure 5 the prediction performance ofSSRM outperforms other two approaches The performanceof IBCF is similar to UBCF As the values of 119896 increase betteraccuracy can be achieved This indicates that more similarneighbor records can provide more information for missingvalue prediction However when 119896 is larger than 25 MAEand RSME fail to drop with the value of 119896 increasing Themain reason is the limited number of similar neighbors Theobservations also show that RMSE has the similar trend butwith larger fluctuations

Next we investigate the impact of matrix density Thematrix density varies from 10 to 50 in increment of10 Similar user size k is set to 5 in this experimentFigure 6 shows the experimental results for response-timeand throughput

As presented in Figure 6 the prediction performance isalso enhanced with the value of matrix density increasingThe main reason is that denser user-service matrix providesmore information for missing value prediction NARM hasthe better performance than LBCF and UBCF under allexperimental setting consistently since NARM considerscontext similarity and the relational degree among cloudservices The experimental results of LBCF and UBCF aresimilar in this experiment since these two similarity compu-tation methods are similar to each other and are both rating-oriented

43 Performance Comparison with Other Popular ApproachesTo study the prediction performance of SSRM we com-pare SSRM with three existing QoS properties predictionapproaches CloudRank2 [3] TECSS [4] and JV-PCC [5]

The size of Top-119896 similar service is an important factorin CF approach which determines how many neighborsrsquohistorical records are employed to generate predictions [5]Therefore the impact of matrix density is not considered in

the experiments We set the density to 10 We vary 119896 from 5to 30 in increment of 5 In order to simplify the experimentalprocess we consider all cloud services are trustworthyFigure 7 shows the experimental results for response-timeand throughput Under the same simulation condition SSRMsignificantly outperforms CloudRank2 TECSS and JV-PCCThe observations also suggest that better accuracy can beachieved by our model when more historical records areavailable in the service selection studyThemain reason is thatcomputing context similarity can improve the performance ofprediction evaluation in MCC environments

Figures 7(a) and 7(b) depict the MAE fractions of SSRMCloudRank2 TECSS and JV-PCC for response-time andthroughput while Figures 7(c) and 7(d) depict the RMSEfractions It can be observed that SSRM achieves smallerMAE and RMSE consistently than other approaches for bothresponse-time and throughput

5 Conclusions and Future

In the paper we propose a novel service selection and recom-mendation model (SSRM) for mobile cloud computing envi-ronmentsWhen calculating user similarity context informa-tion and interest are considered Through utilizing relationaldegree to improve the accuracy of ranking result our serviceselection and recommendation method succeed in obtainingthe ultimate service ranking results In addition SSRM onlyselects trusted mobile cloud service as candidates Thereforea set of trusted candidates are identified for target user Theexperimental results show that our approach significantlyimproves the prediction performance as comparedwith othertwo traditional approaches item-based CF approach (IBCF)and user-based CF approach (UBCF)

There are some disadvantages in our SSRM Firstly weexploit node distance to compute context similarity Thereis an underlying assumption in this exploitation each ofthe two adjacent nodes has equal semantic distance orgranularity of nodes in each level is identicalThis underlyingassumption is not true in some scenarios and it deterioratesthe performance of similarity evaluation Secondly the trustthreshold is used to select trusted mobile cloud service ascandidates However it cannot detect and exclude maliciousQoS values provided by users Thirdly in our experimentsthere are only three types of context information that areconsidered

Therefore in the future we will study the methods toimprove context similarity measurement We will study how

Wireless Communications and Mobile Computing 11

Response-time

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E10 25 3015 205

k

(b)Response-time

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

(c)

Throughput

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

(d)

Figure 5 Impact of neighbor size 119896 comparison with SSRM IBCF and UBCF

12 Wireless Communications and Mobile Computing

Response-time

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E

20 50 6030 4010Matrix density ()

(a)

Throughput

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E20 30 40 50 6010

Matrix density ()

(b)

Response-time

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 50 6030 4010Matrix density ()

(c)

Throughput

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 30 40 50 6010Matrix density ()

(d)

Figure 6 Impact of matrix density comparison with SSRM IBCF and UBCF

Wireless Communications and Mobile Computing 13

Response-time

JV-PCCTECSSSSRM

CloudRank2

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

02

025

03

035

04

045

05

055

06

065

07

MA

E10 15 20 25 305

k

JV-PCCTECSSSSRM

CloudRank2

(b)

Response-time

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

JV-PCCTECSSSSRM

CloudRank2

(c)

Throughput

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

JV-PCCTECSSSSRM

CloudRank2

(d)

Figure 7 Impact of neighbor size 119896 comparison with other popular approaches

14 Wireless Communications and Mobile Computing

to select themost trustworthy cloud service of certain type forthe active users based on multidimensional trust evidenceWe also will conduct more experimental investigations thatdeal with the impact of different context changes on serviceselection Moreover we will improve the ranking accuracyof our approaches by exploiting additional techniques (com-bining rating-oriented approaches and ranking-orientedapproaches matrix factorization intelligent optimizationalgorithms etc)

Conflicts of Interest

The author declares that they have no conflicts of interest

Acknowledgments

The work in this paper has been supported by NationalNatural Science Foundation of China (Program no 71501156)and China Postdoctoral Science Foundation (Program no2014M560796) and Shanxi Provincial EducationDepartment(Program no 15JK1679)

References

[1] White PaperMobile CloudComputing Solution Brief AEPONA2010

[2] httpsenwikipediaorgwikiCloudlet[3] Z Zheng X Wu Y Zhang M R Lyu and J Wang ldquoQoS

ranking prediction for cloud servicesrdquo IEEE Transactions onParallel and Distributed Systems vol 24 no 6 pp 1213ndash12222013

[4] Y Pan SDingW Fan J Li and S Yang ldquoTrust-enhanced cloudservice selection model based on QoS analysisrdquo PLoS ONE vol10 no 11 Article ID e0143448 2015

[5] S Ding C-Y Xia K-L Zhou S-L Yang and J S Shang ldquoDeci-sion support for personalized cloud service selection throughmulti-attribute trustworthiness evaluationrdquo PLoS ONE vol 9no 6 Article ID e107821 2014

[6] S K Garg S Versteeg and R Buyya ldquoA framework for rankingof cloud computing servicesrdquo Future Generation ComputerSystems vol 29 no 4 pp 1012ndash1023 2013

[7] H Ma Z Hu K Li and H Zhang ldquoToward trustworthy cloudservice selection a time-aware approach using interval neutro-sophic setrdquo Journal of Parallel and Distributed Computing vol96 pp 75ndash94 2016

[8] A Ahmed and A S Sabyasachi ldquoMobile cloud service selectionusing back propagation neural networkrdquo International Journalof Computer Applications vol 129 no 14 pp 1ndash5 2015

[9] L Guo J Ma Z-M Chen and H-R Jiang ldquoIncorporatingitem relations for social recommendationrdquo Chinese Journal ofComputers vol 37 no 1 pp 219ndash228 2014

[10] Z Yang B Wu K Zheng X Wang and L Lei ldquoA survey ofcollaborative filtering-based recommender systems for mobileinternet applicationsrdquo IEEE Access vol 4 pp 3273ndash3287 2016

[11] Markets and Markets httpwwwmarketsandmarketscomPressReleasesmobile-cloudasp

[12] M Whaiduzzaman A Gani N B Anuar M Shiraz MN Haque and I T Haque ldquoCloud service selection usingmulticriteria decision analysisrdquoTheScientificWorld Journal vol2014 Article ID 459375 10 pages 2014

[13] K Tserpes F Aisopos D Kyriazis and T Varvarigou ldquoArecommender mechanism for service selection in service-oriented environmentsrdquo Future Generation Computer Systemsvol 28 no 8 pp 1285ndash1294 2012

[14] Z U Rehman O K Hussain and F K Hussain ldquoParallelcloud service selection and ranking based on QoS historyrdquoInternational Journal of Parallel Programming vol 42 no 5 pp820ndash852 2014

[15] W-J Fan S-L YangH Perros and J Pei ldquoAmulti-dimensionaltrust-aware cloud service selection mechanism based on evi-dential reasoning approachrdquo International Journal of Automa-tion and Computing vol 12 no 2 pp 208ndash219 2015

[16] X G Wang J Cao and Y Xiang ldquoDynamic cloud serviceselection using an adaptive learning mechanism in multi-cloudcomputingrdquo The Journal of Systems and Software vol 100 pp195ndash210 2015

[17] C T Do N H Tran E-N Huh C S Hong D Niyato andZ Han ldquoDynamics of service selection and provider pricinggame in heterogeneous cloud marketrdquo Journal of Network andComputer Applications vol 69 pp 152ndash165 2015

[18] R Pal and P Hui ldquoEconomic models for cloud service marketspricing and capacity planningrdquo Theoretical Computer Sciencevol 496 pp 113ndash124 2013

[19] A Ezenwoke O Daramola and M Adigun ldquoTowards avisualization framework for service selection in cloud e-marketplacesrdquo in Proceedings of the 2017 IEEE InternationalConference on Web Services (ICWS rsquo17) pp 828ndash835 HonoluluHI USA June 2017

[20] M Tang X Dai J Liu and J Chen ldquoTowards a trust evaluationmiddleware for cloud service selectionrdquo Future GenerationComputer Systems vol 74 pp 302ndash312 2017

[21] A K Dey ldquoUnderstanding and using contextrdquo Personal andUbiquitous Computing vol 5 no 1 pp 4ndash7 2001

[22] X Han L Wang S Park A Cuevas and N Crespi ldquoAlikepeople alike interests A large-scale study on interest similarityin social networksrdquo in Proceedings of the 2014 IEEEACM Inter-national Conference on Advances in Social Networks Analysisand Mining (ASONAM rsquo14) pp 491ndash496 August 2014

[23] R N Lichtenwalter J T Lussier and N V Chawla ldquoNewperspectives and methods in link predictionrdquo in Proceedings ofthe 16th ACM SIGKDD International Conference on KnowledgeDiscovery and Data Mining (KDD rsquo10) pp 243ndash252 July 2010

[24] L Munasinghe and R Ichise ldquoLink prediction in social net-works using information flow via active linksrdquo IEICE Transac-tion on Information and Systems vol E96-D no 7 pp 1495ndash1502 2013

[25] Axis httpaxisapacheorgaxis2[26] ns3 httpswwwnsnamorgns3[27] Dataset httpwwwzibinzhengcomtpds

Wireless Communications and Mobile Computing 7

Inputan employed service set 119864a full service set 119868an employed similar training user set 119873119862mcs119895

(119906119909)observed service usage experience values 119903119906119910 mcs119895 (119906119910 isin 119873119862mcs119895

(119906119909))Outa service ranking 119865 = 119864while 119865 = Oslash do119905 = argmaxmcs119895isin119865119876119906119909 mcs119895 119877119890(119905) = |119864| minus |119865| + 1119865 = 119865 minus 119905endforeach mcs119895 isin 119868 do

119876119906119909 mcs119895 =

119903119906119909 mcs119895 minus min(119903119906119909 )max(119903119906119909 ) minus min(119903119906119909 ) 119903119906119909 mcs119895 isin ldquobenefitrdquo

max(119903119906119909 ) minus 119903119906119909 mcs119895

max(119903119906119909 ) minus min(119903119906119909 ) 119903119906119909 mcs119895 isin ldquocostrdquo

end119899 = |119868|while 119868 = Oslash do119905 = argmaxmcs119895isin119868119876119906119909 mcs119895(119905) = 119899 minus |119868| + 1119868 = 119868 minus 119905end

end

Algorithm 1 Basic ranking algorithm

32 Improving Accuracy of Ranking Results Through calcu-lating the similar relation among users the missing value119903119906119909mcs119895 is successfully predicted in Section 31 However therelation among cloud services is no doubt an importantevaluation factor of missing value prediction The relatedservices will be probably selected by the same user based onthe intuition [9]Therefore we calculate the relational degreeand utilize it to improve the accuracy of missing value 119903119906119909mcs119895computation in the section The relational degree betweenservice mcs119895 and service mcs119894 is denoted as 119889119895119894 In the paperPropFlow [23] is used to calculate 119889119895119894 PropFlow algorithmcomputes the information flow between services where alarger value indicates tighter relation In order to calculate119889119895119894 we firstly describe the formal definition of service relationgraph

Definition 2 (service relation graph) Given a set of servicesMCS and a totally ordered domain of weights 119882 a servicerelation graph is a weighted undirected graph 119866 = (MCS 119864)The edge (mcs119894mcs119895 119908119894119895) in the set 119864 isin 119880 times 119880 times 119882 encodesthe link weight119908119894119895 (119908119894119895 ge 0) between service mcs119894 and servicemcs119895 An edge depicts the direct link relationship betweenmcs119894 and mcs119895 As 119866 is undirected 119908119894119895 = 119908119895119894

Equation (10) shows how to calculate 119889119895119894119889119895119894 = Input119895 lowast 119908119895119894sum119896isin119873(119895) 119908119895119896 (10)

Ｇ＝Ｍℎ

Ｇ＝ＭiＧ＝Ｍl

Ｇ＝Ｍe

Ｇ＝Ｍf

Ｇ＝Ｍj Ｇ＝Ｍk

wkl = 1 wli = 5

wle = 1

wie = 1wke = 1

wkℎ = 2

wjk = 3

wjf = 1

Figure 3 An example of service relation graph

where initial input is regarded as 1 and 119873(119895) is the set ofneighbors of mcs119895

Figure 3 shows an example of relation degree computa-tion where there are six kinds of mobile cloud services mcs119894and mcs119895 are indirectly linked We assume mcs119895 is startingnode

Four paths can reach mcs119894 from mcs119895 but only twopaths are considered for computation based on PropFlow

8 Wireless Communications and Mobile Computing

algorithm They are mcs119895 rarr mcs119896 rarr mcs119897 rarr mcs119894 andmcs119895 rarr mcs119896 rarr mcs119890 rarr mcs119894

First we compute 119889119895119896 The sum of link weights of mcs119895and its neighbor is 4 119889119895119896 is computed as

119889119895119896 = 1 lowast 3(1 + 3) = 1 lowast 34 = 34 (11)

119889119896119897 is computed as

119889119896119897 = 119889119895119896 lowast 1(1 + 1 + 2) = 34 lowast 14 = 316 (12)

With the same method 119889119895119894 is computed as

119889119895119894 = 119889119897119894 + 119889119890119894 = 119889119896119897 lowast 55 + 119889119896119890 lowast 11 = 38 (13)

More details of calculating 119889119895119894 are shown in [24]Next we assume service mcs119894 is invoked recently by user119906119909 119889119895119894 is incorporated into (7) and then the missing value119903119906119909mcs119895 is computed with (14) in basic ranking algorithm

Finally the ultimate ranking results are got

119903119906119909 mcs119895 = (119903119906119909+ sum119906119910isin119873119862mcs119895

(119906119909)(119903119906119910 mcs119895 minus 119903119906119910) sim119880 (119906119909 119906119910)sum119906119910isin119873119862mcs119895(119906119909)

sim119880 (119906119909 119906119910) )sdot (1 + 119889119895119894)

(14)

In SSRM we only select trusted mobile cloud serviceas candidates Therefore a set of trusted candidates areidentified for 119906119909 by

MCS119909 = mcs119895 | trustmcs119895 gt 120583119909 119895 = 1 119873 (15)

where trustmcs119895 denotes the trustworthiness of mobile cloudservice mcs119895 and 120583119909 expresses the trust threshold decidedby user 119906119909 Many existing methods [4ndash6] can be applied tocompute trustmcs119895 Here how to compute trustmcs119895 is not tobe discussed We omit the details for brevity

4 Experimental Study

In this section in order to evaluate the effectiveness ofSSRM a series of test scenarios are developed To studythe prediction performance we compare SSRM with othertwo traditional approaches item-based CF approach (IBCF)and user-based CF approach (UBCF) SSRM IBCF andUBCF are rating-oriented methods which rank the cloudservices based on the predicted QoS values Since there isno suitable real data supporting mobile cloud computingsimulation we use ns3 simulator to generate the experimentaldataset Firstly we generated cloud service invocation codes

by Axis2 [25] a Java-based open source package for cloudservices Then we simulate mobile cloud service on ns3[26] based on the invocation codes On ns3 the creationof servers mobile users and service model was easier thanwith other network simulators In the simulating process ofdataset the real-world web service QoS dataset from WS-DREAM team [3] is referenced by usThe detailed real-worldQoS values are publicly released online [27] which makesour experimental evaluations reproducible We simulate 100mobile users 25 servers and 300 services The value of linkweight between two services is randomly generated which isin the range of [1 5] In order to conduct our experimentsrealistically the changes of user context information aresimulated Three types of context are considered includingbandwidth memory and battery capacity The bandwidthvalue of a mobile user is in the range of 4Mndash8MThe valuesof memory and battery capacity will decline over timeThesevalues are in the range of 100000000ndash2100000000MB and05ndash2KVA

We normalize the context information by

119862nor119906119909mcs119895 (V) = 119886 tan (119862119906119909 mcs119895 (V)) lowast 2120587 (16)

where 119862nor119906119909mcs119895(V) denotes the normalized context value of

user 119906119909 over cloud service mcs119895 Inverse tangent function isapplied to normalize the raw data 119862119906119909 mcs119895(V) Then thesenormalized property values are used to calculate contextsimilarity with (4) Once the context simulation is finishedthe target user 119906119909 will obtain context similarity results foreach cloud service

Similar to [3] QoS properties include throughput andresponse-time that are in the range of 0ndash1000 kbps and 0ndash20second respectively Most of the throughput and response-time values fall between 5ndash40 kbps and 01ndash08 seconds Inorder to simply the experimental process 200 service QoSrecords are randomly selected and two 100 times 200 user-servicematrices are constructed The two matrices include through-put and response-time respectivelyUser-servicematrices areseparated into two parts training set (80 historical usageexperiences in the matrix) and test set (the remaining 20usage experiences) Each entry in the matrix is the QoSvalue (eg response-time or throughput) of a mobile cloudservice observed by a user The experiments employ the QoSvalues of response-time and throughput to rank the servicesindependently Table 1 shows descriptions of the obtainedpractical cloud service QoS values and context information

We use Mean Absolute Error (MAE) and Root MeanSquare Error (RMSE) as the metric to evaluate predictionperformance of the proposed approach in comparison withother approaches MAE and RMSE are defined as

MAE = sum119906119909 mcs119895

100381610038161003816100381610038161003816119903act119906119909mcs119895 minus 119903pre119906119909mcs119895100381610038161003816100381610038161003816119897pre

RMSE = radic sum119906119909 mcs119895 (119903act119906119909mcs119895 minus 119903pre119906119909 mcs119895)2119897pre (17)

Wireless Communications and Mobile Computing 9

Response-time

02

025

03

035

04

045

05

055

06

065

07

MA

E

01 02 03 04 05 06 07 08 09 100Weighting factor

(a) MAE values

Response-time

03

035

04

045

05

055

06

065

07

075

08

RMSE

01 02 03 04 05 06 07 08 09 100Weighting factor

(b) RMSE values

Figure 4 Impact of weighting factor 120582

Table 1 Mobile cloud service QoS dataset and context informationdescriptions

Statistics ValueNumber of mobile cloud serviceinvocations 180000

Number of service users 100Number of mobile cloud services 200Minimum response-time value 0004 sMaximum response-time value 20 sMean of response-time 110 sStandard deviation ofresponse-time 226 s

Minimum throughput value 01 kbpsMaximum throughput value 1000 kbpsMean of throughput 2546 kbpsStandard deviation of throughput 4803 kbpsBandwidth 4Mndash8MMemory 100000000MBndash2100000000MBBattery capacity 05 KVAndash2KVA

where 119903act119906119909mcs119895 and 119903pre119906119909mcs119895 denote the actual QoS valueand predicted value respectively 119897pre is the number ofpredicted QoS values Smaller values of MAE and RMSEindicate better results

41 Impact of 120582 120582 is weighting factor in (2) 120582 determineshow much similarity measure relies on context and interestIn the experiment we vary the weighting factor 120582 from 0 to 1in increment of 01 The Top-k is set to 5 Matrix density is setto 10 Matrix density means the percentage of selected QoSentries that are used to predict the missing QoS value Theexperimental results are shown in Figure 4 As the value of 120582increases MAE firstly decreases and then quickly increasesWhen 120582 = 04 our results have the best value As shown inFigure 4(b) RMSE presents similar trend

42 Performance Comparison of SSRM IBCF and UBCFTo compare the performance of SSRM IBCF and UBCFsimilarity computation methods we implement IBCF andUBCF methods

Item-Based CF Approach (IBCF) We apply PCC to calculatesimilarities between users and predicted QoS values based onsimilar users The user similarity is calculated by

sim (119906119909 119906119910) = summcs119895isinmcs119906119909119906119910(119903119906119909 mcs119895 minus 119903119906119909) (119903119906119910 mcs119895 minus 119903119906119910)

radicsummcs119895isinmcs119906119909119906119910(119903119906119909mcs119895 minus 119903119906119909)2radicsummcs119895isinmcs119906119909119906119910

(119903119906119910mcs119895 minus 119903119906119910)2 (18)

10 Wireless Communications and Mobile Computing

where mcs119906119909119906119910 is the set of mobile cloud services that havebeen coinvoked by 119906119909 and 119906119910 119903119906119909and 119903119906119910 are average QoSvalues of cloud service invoked by 119906119909 and 119906119910 respectively

User-Based CF Approach (UBCF) We apply PCC to calculatesimilarities between services and predicted QoS values basedon similar services The service similarity is calculated by

sim (mcs119895mcs119894) = sum119906119909isin119906mcs119895mcs119894(119903119906119909mcs119895 minus 119903mcs119895) (119903119906119909 mcs119894 minus 119903mcs119894)

radicsum119906119909isin119906mcs119895mcs119894(119903119906119909mcs119895 minus 119903mcs119895)2radicsum119906119909isin119906mcs119895mcs119894

(119903119906119909mcs119894 minus 119903mcs119894)2 (19)

where 119906mcs119895mcs119894 is the set of users who have invoked the cloudservices mcs119895 andmcs119894 119903mcs119895 and 119903mcs119894 are average QoS valuesof cloud services mcs119895 and mcs119894 made by 119906119909

In the experiments trustworthiness of cloud service hasno influence on similarity computation and service decisionin the experiments In order to simplify the experimentalprocess we consider all cloud services are trustworthyWeighting factor 120582 is set to 04 We firstly study the effectof neighbor size 119896 Matrix density is set to 10 We change119896 from 5 to 30 in increment of 5 Figure 5 shows theexperimental results for response-time and throughput

As shown in Figure 5 the prediction performance ofSSRM outperforms other two approaches The performanceof IBCF is similar to UBCF As the values of 119896 increase betteraccuracy can be achieved This indicates that more similarneighbor records can provide more information for missingvalue prediction However when 119896 is larger than 25 MAEand RSME fail to drop with the value of 119896 increasing Themain reason is the limited number of similar neighbors Theobservations also show that RMSE has the similar trend butwith larger fluctuations

Next we investigate the impact of matrix density Thematrix density varies from 10 to 50 in increment of10 Similar user size k is set to 5 in this experimentFigure 6 shows the experimental results for response-timeand throughput

As presented in Figure 6 the prediction performance isalso enhanced with the value of matrix density increasingThe main reason is that denser user-service matrix providesmore information for missing value prediction NARM hasthe better performance than LBCF and UBCF under allexperimental setting consistently since NARM considerscontext similarity and the relational degree among cloudservices The experimental results of LBCF and UBCF aresimilar in this experiment since these two similarity compu-tation methods are similar to each other and are both rating-oriented

43 Performance Comparison with Other Popular ApproachesTo study the prediction performance of SSRM we com-pare SSRM with three existing QoS properties predictionapproaches CloudRank2 [3] TECSS [4] and JV-PCC [5]

The size of Top-119896 similar service is an important factorin CF approach which determines how many neighborsrsquohistorical records are employed to generate predictions [5]Therefore the impact of matrix density is not considered in

the experiments We set the density to 10 We vary 119896 from 5to 30 in increment of 5 In order to simplify the experimentalprocess we consider all cloud services are trustworthyFigure 7 shows the experimental results for response-timeand throughput Under the same simulation condition SSRMsignificantly outperforms CloudRank2 TECSS and JV-PCCThe observations also suggest that better accuracy can beachieved by our model when more historical records areavailable in the service selection studyThemain reason is thatcomputing context similarity can improve the performance ofprediction evaluation in MCC environments

Figures 7(a) and 7(b) depict the MAE fractions of SSRMCloudRank2 TECSS and JV-PCC for response-time andthroughput while Figures 7(c) and 7(d) depict the RMSEfractions It can be observed that SSRM achieves smallerMAE and RMSE consistently than other approaches for bothresponse-time and throughput

5 Conclusions and Future

In the paper we propose a novel service selection and recom-mendation model (SSRM) for mobile cloud computing envi-ronmentsWhen calculating user similarity context informa-tion and interest are considered Through utilizing relationaldegree to improve the accuracy of ranking result our serviceselection and recommendation method succeed in obtainingthe ultimate service ranking results In addition SSRM onlyselects trusted mobile cloud service as candidates Thereforea set of trusted candidates are identified for target user Theexperimental results show that our approach significantlyimproves the prediction performance as comparedwith othertwo traditional approaches item-based CF approach (IBCF)and user-based CF approach (UBCF)

There are some disadvantages in our SSRM Firstly weexploit node distance to compute context similarity Thereis an underlying assumption in this exploitation each ofthe two adjacent nodes has equal semantic distance orgranularity of nodes in each level is identicalThis underlyingassumption is not true in some scenarios and it deterioratesthe performance of similarity evaluation Secondly the trustthreshold is used to select trusted mobile cloud service ascandidates However it cannot detect and exclude maliciousQoS values provided by users Thirdly in our experimentsthere are only three types of context information that areconsidered

Therefore in the future we will study the methods toimprove context similarity measurement We will study how

Wireless Communications and Mobile Computing 11

Response-time

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E10 25 3015 205

k

(b)Response-time

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

(c)

Throughput

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

(d)

Figure 5 Impact of neighbor size 119896 comparison with SSRM IBCF and UBCF

12 Wireless Communications and Mobile Computing

Response-time

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E

20 50 6030 4010Matrix density ()

(a)

Throughput

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E20 30 40 50 6010

Matrix density ()

(b)

Response-time

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 50 6030 4010Matrix density ()

(c)

Throughput

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 30 40 50 6010Matrix density ()

(d)

Figure 6 Impact of matrix density comparison with SSRM IBCF and UBCF

Wireless Communications and Mobile Computing 13

Response-time

JV-PCCTECSSSSRM

CloudRank2

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

02

025

03

035

04

045

05

055

06

065

07

MA

E10 15 20 25 305

k

JV-PCCTECSSSSRM

CloudRank2

(b)

Response-time

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

JV-PCCTECSSSSRM

CloudRank2

(c)

Throughput

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

JV-PCCTECSSSSRM

CloudRank2

(d)

Figure 7 Impact of neighbor size 119896 comparison with other popular approaches

14 Wireless Communications and Mobile Computing

to select themost trustworthy cloud service of certain type forthe active users based on multidimensional trust evidenceWe also will conduct more experimental investigations thatdeal with the impact of different context changes on serviceselection Moreover we will improve the ranking accuracyof our approaches by exploiting additional techniques (com-bining rating-oriented approaches and ranking-orientedapproaches matrix factorization intelligent optimizationalgorithms etc)

Conflicts of Interest

The author declares that they have no conflicts of interest

Acknowledgments

The work in this paper has been supported by NationalNatural Science Foundation of China (Program no 71501156)and China Postdoctoral Science Foundation (Program no2014M560796) and Shanxi Provincial EducationDepartment(Program no 15JK1679)

References

[1] White PaperMobile CloudComputing Solution Brief AEPONA2010

[2] httpsenwikipediaorgwikiCloudlet[3] Z Zheng X Wu Y Zhang M R Lyu and J Wang ldquoQoS

ranking prediction for cloud servicesrdquo IEEE Transactions onParallel and Distributed Systems vol 24 no 6 pp 1213ndash12222013

[4] Y Pan SDingW Fan J Li and S Yang ldquoTrust-enhanced cloudservice selection model based on QoS analysisrdquo PLoS ONE vol10 no 11 Article ID e0143448 2015

[5] S Ding C-Y Xia K-L Zhou S-L Yang and J S Shang ldquoDeci-sion support for personalized cloud service selection throughmulti-attribute trustworthiness evaluationrdquo PLoS ONE vol 9no 6 Article ID e107821 2014

[6] S K Garg S Versteeg and R Buyya ldquoA framework for rankingof cloud computing servicesrdquo Future Generation ComputerSystems vol 29 no 4 pp 1012ndash1023 2013

[7] H Ma Z Hu K Li and H Zhang ldquoToward trustworthy cloudservice selection a time-aware approach using interval neutro-sophic setrdquo Journal of Parallel and Distributed Computing vol96 pp 75ndash94 2016

[8] A Ahmed and A S Sabyasachi ldquoMobile cloud service selectionusing back propagation neural networkrdquo International Journalof Computer Applications vol 129 no 14 pp 1ndash5 2015

[9] L Guo J Ma Z-M Chen and H-R Jiang ldquoIncorporatingitem relations for social recommendationrdquo Chinese Journal ofComputers vol 37 no 1 pp 219ndash228 2014

[10] Z Yang B Wu K Zheng X Wang and L Lei ldquoA survey ofcollaborative filtering-based recommender systems for mobileinternet applicationsrdquo IEEE Access vol 4 pp 3273ndash3287 2016

[11] Markets and Markets httpwwwmarketsandmarketscomPressReleasesmobile-cloudasp

[12] M Whaiduzzaman A Gani N B Anuar M Shiraz MN Haque and I T Haque ldquoCloud service selection usingmulticriteria decision analysisrdquoTheScientificWorld Journal vol2014 Article ID 459375 10 pages 2014

[13] K Tserpes F Aisopos D Kyriazis and T Varvarigou ldquoArecommender mechanism for service selection in service-oriented environmentsrdquo Future Generation Computer Systemsvol 28 no 8 pp 1285ndash1294 2012

[14] Z U Rehman O K Hussain and F K Hussain ldquoParallelcloud service selection and ranking based on QoS historyrdquoInternational Journal of Parallel Programming vol 42 no 5 pp820ndash852 2014

[15] W-J Fan S-L YangH Perros and J Pei ldquoAmulti-dimensionaltrust-aware cloud service selection mechanism based on evi-dential reasoning approachrdquo International Journal of Automa-tion and Computing vol 12 no 2 pp 208ndash219 2015

[16] X G Wang J Cao and Y Xiang ldquoDynamic cloud serviceselection using an adaptive learning mechanism in multi-cloudcomputingrdquo The Journal of Systems and Software vol 100 pp195ndash210 2015

[17] C T Do N H Tran E-N Huh C S Hong D Niyato andZ Han ldquoDynamics of service selection and provider pricinggame in heterogeneous cloud marketrdquo Journal of Network andComputer Applications vol 69 pp 152ndash165 2015

[18] R Pal and P Hui ldquoEconomic models for cloud service marketspricing and capacity planningrdquo Theoretical Computer Sciencevol 496 pp 113ndash124 2013

[19] A Ezenwoke O Daramola and M Adigun ldquoTowards avisualization framework for service selection in cloud e-marketplacesrdquo in Proceedings of the 2017 IEEE InternationalConference on Web Services (ICWS rsquo17) pp 828ndash835 HonoluluHI USA June 2017

[20] M Tang X Dai J Liu and J Chen ldquoTowards a trust evaluationmiddleware for cloud service selectionrdquo Future GenerationComputer Systems vol 74 pp 302ndash312 2017

[21] A K Dey ldquoUnderstanding and using contextrdquo Personal andUbiquitous Computing vol 5 no 1 pp 4ndash7 2001

[22] X Han L Wang S Park A Cuevas and N Crespi ldquoAlikepeople alike interests A large-scale study on interest similarityin social networksrdquo in Proceedings of the 2014 IEEEACM Inter-national Conference on Advances in Social Networks Analysisand Mining (ASONAM rsquo14) pp 491ndash496 August 2014

[23] R N Lichtenwalter J T Lussier and N V Chawla ldquoNewperspectives and methods in link predictionrdquo in Proceedings ofthe 16th ACM SIGKDD International Conference on KnowledgeDiscovery and Data Mining (KDD rsquo10) pp 243ndash252 July 2010

[24] L Munasinghe and R Ichise ldquoLink prediction in social net-works using information flow via active linksrdquo IEICE Transac-tion on Information and Systems vol E96-D no 7 pp 1495ndash1502 2013

[25] Axis httpaxisapacheorgaxis2[26] ns3 httpswwwnsnamorgns3[27] Dataset httpwwwzibinzhengcomtpds

8 Wireless Communications and Mobile Computing

algorithm They are mcs119895 rarr mcs119896 rarr mcs119897 rarr mcs119894 andmcs119895 rarr mcs119896 rarr mcs119890 rarr mcs119894

First we compute 119889119895119896 The sum of link weights of mcs119895and its neighbor is 4 119889119895119896 is computed as

119889119895119896 = 1 lowast 3(1 + 3) = 1 lowast 34 = 34 (11)

119889119896119897 is computed as

119889119896119897 = 119889119895119896 lowast 1(1 + 1 + 2) = 34 lowast 14 = 316 (12)

With the same method 119889119895119894 is computed as

119889119895119894 = 119889119897119894 + 119889119890119894 = 119889119896119897 lowast 55 + 119889119896119890 lowast 11 = 38 (13)

More details of calculating 119889119895119894 are shown in [24]Next we assume service mcs119894 is invoked recently by user119906119909 119889119895119894 is incorporated into (7) and then the missing value119903119906119909mcs119895 is computed with (14) in basic ranking algorithm

Finally the ultimate ranking results are got

119903119906119909 mcs119895 = (119903119906119909+ sum119906119910isin119873119862mcs119895

(119906119909)(119903119906119910 mcs119895 minus 119903119906119910) sim119880 (119906119909 119906119910)sum119906119910isin119873119862mcs119895(119906119909)

sim119880 (119906119909 119906119910) )sdot (1 + 119889119895119894)

(14)

In SSRM we only select trusted mobile cloud serviceas candidates Therefore a set of trusted candidates areidentified for 119906119909 by

MCS119909 = mcs119895 | trustmcs119895 gt 120583119909 119895 = 1 119873 (15)

where trustmcs119895 denotes the trustworthiness of mobile cloudservice mcs119895 and 120583119909 expresses the trust threshold decidedby user 119906119909 Many existing methods [4ndash6] can be applied tocompute trustmcs119895 Here how to compute trustmcs119895 is not tobe discussed We omit the details for brevity

4 Experimental Study

In this section in order to evaluate the effectiveness ofSSRM a series of test scenarios are developed To studythe prediction performance we compare SSRM with othertwo traditional approaches item-based CF approach (IBCF)and user-based CF approach (UBCF) SSRM IBCF andUBCF are rating-oriented methods which rank the cloudservices based on the predicted QoS values Since there isno suitable real data supporting mobile cloud computingsimulation we use ns3 simulator to generate the experimentaldataset Firstly we generated cloud service invocation codes

by Axis2 [25] a Java-based open source package for cloudservices Then we simulate mobile cloud service on ns3[26] based on the invocation codes On ns3 the creationof servers mobile users and service model was easier thanwith other network simulators In the simulating process ofdataset the real-world web service QoS dataset from WS-DREAM team [3] is referenced by usThe detailed real-worldQoS values are publicly released online [27] which makesour experimental evaluations reproducible We simulate 100mobile users 25 servers and 300 services The value of linkweight between two services is randomly generated which isin the range of [1 5] In order to conduct our experimentsrealistically the changes of user context information aresimulated Three types of context are considered includingbandwidth memory and battery capacity The bandwidthvalue of a mobile user is in the range of 4Mndash8MThe valuesof memory and battery capacity will decline over timeThesevalues are in the range of 100000000ndash2100000000MB and05ndash2KVA

We normalize the context information by

119862nor119906119909mcs119895 (V) = 119886 tan (119862119906119909 mcs119895 (V)) lowast 2120587 (16)

where 119862nor119906119909mcs119895(V) denotes the normalized context value of

user 119906119909 over cloud service mcs119895 Inverse tangent function isapplied to normalize the raw data 119862119906119909 mcs119895(V) Then thesenormalized property values are used to calculate contextsimilarity with (4) Once the context simulation is finishedthe target user 119906119909 will obtain context similarity results foreach cloud service

Similar to [3] QoS properties include throughput andresponse-time that are in the range of 0ndash1000 kbps and 0ndash20second respectively Most of the throughput and response-time values fall between 5ndash40 kbps and 01ndash08 seconds Inorder to simply the experimental process 200 service QoSrecords are randomly selected and two 100 times 200 user-servicematrices are constructed The two matrices include through-put and response-time respectivelyUser-servicematrices areseparated into two parts training set (80 historical usageexperiences in the matrix) and test set (the remaining 20usage experiences) Each entry in the matrix is the QoSvalue (eg response-time or throughput) of a mobile cloudservice observed by a user The experiments employ the QoSvalues of response-time and throughput to rank the servicesindependently Table 1 shows descriptions of the obtainedpractical cloud service QoS values and context information

We use Mean Absolute Error (MAE) and Root MeanSquare Error (RMSE) as the metric to evaluate predictionperformance of the proposed approach in comparison withother approaches MAE and RMSE are defined as

MAE = sum119906119909 mcs119895

100381610038161003816100381610038161003816119903act119906119909mcs119895 minus 119903pre119906119909mcs119895100381610038161003816100381610038161003816119897pre

RMSE = radic sum119906119909 mcs119895 (119903act119906119909mcs119895 minus 119903pre119906119909 mcs119895)2119897pre (17)

Wireless Communications and Mobile Computing 9

Response-time

02

025

03

035

04

045

05

055

06

065

07

MA

E

01 02 03 04 05 06 07 08 09 100Weighting factor

(a) MAE values

Response-time

03

035

04

045

05

055

06

065

07

075

08

RMSE

01 02 03 04 05 06 07 08 09 100Weighting factor

(b) RMSE values

Figure 4 Impact of weighting factor 120582

Table 1 Mobile cloud service QoS dataset and context informationdescriptions

Statistics ValueNumber of mobile cloud serviceinvocations 180000

Number of service users 100Number of mobile cloud services 200Minimum response-time value 0004 sMaximum response-time value 20 sMean of response-time 110 sStandard deviation ofresponse-time 226 s

Minimum throughput value 01 kbpsMaximum throughput value 1000 kbpsMean of throughput 2546 kbpsStandard deviation of throughput 4803 kbpsBandwidth 4Mndash8MMemory 100000000MBndash2100000000MBBattery capacity 05 KVAndash2KVA

where 119903act119906119909mcs119895 and 119903pre119906119909mcs119895 denote the actual QoS valueand predicted value respectively 119897pre is the number ofpredicted QoS values Smaller values of MAE and RMSEindicate better results

41 Impact of 120582 120582 is weighting factor in (2) 120582 determineshow much similarity measure relies on context and interestIn the experiment we vary the weighting factor 120582 from 0 to 1in increment of 01 The Top-k is set to 5 Matrix density is setto 10 Matrix density means the percentage of selected QoSentries that are used to predict the missing QoS value Theexperimental results are shown in Figure 4 As the value of 120582increases MAE firstly decreases and then quickly increasesWhen 120582 = 04 our results have the best value As shown inFigure 4(b) RMSE presents similar trend

42 Performance Comparison of SSRM IBCF and UBCFTo compare the performance of SSRM IBCF and UBCFsimilarity computation methods we implement IBCF andUBCF methods

Item-Based CF Approach (IBCF) We apply PCC to calculatesimilarities between users and predicted QoS values based onsimilar users The user similarity is calculated by

sim (119906119909 119906119910) = summcs119895isinmcs119906119909119906119910(119903119906119909 mcs119895 minus 119903119906119909) (119903119906119910 mcs119895 minus 119903119906119910)

radicsummcs119895isinmcs119906119909119906119910(119903119906119909mcs119895 minus 119903119906119909)2radicsummcs119895isinmcs119906119909119906119910

(119903119906119910mcs119895 minus 119903119906119910)2 (18)

10 Wireless Communications and Mobile Computing

where mcs119906119909119906119910 is the set of mobile cloud services that havebeen coinvoked by 119906119909 and 119906119910 119903119906119909and 119903119906119910 are average QoSvalues of cloud service invoked by 119906119909 and 119906119910 respectively

User-Based CF Approach (UBCF) We apply PCC to calculatesimilarities between services and predicted QoS values basedon similar services The service similarity is calculated by

sim (mcs119895mcs119894) = sum119906119909isin119906mcs119895mcs119894(119903119906119909mcs119895 minus 119903mcs119895) (119903119906119909 mcs119894 minus 119903mcs119894)

radicsum119906119909isin119906mcs119895mcs119894(119903119906119909mcs119895 minus 119903mcs119895)2radicsum119906119909isin119906mcs119895mcs119894

(119903119906119909mcs119894 minus 119903mcs119894)2 (19)

where 119906mcs119895mcs119894 is the set of users who have invoked the cloudservices mcs119895 andmcs119894 119903mcs119895 and 119903mcs119894 are average QoS valuesof cloud services mcs119895 and mcs119894 made by 119906119909

In the experiments trustworthiness of cloud service hasno influence on similarity computation and service decisionin the experiments In order to simplify the experimentalprocess we consider all cloud services are trustworthyWeighting factor 120582 is set to 04 We firstly study the effectof neighbor size 119896 Matrix density is set to 10 We change119896 from 5 to 30 in increment of 5 Figure 5 shows theexperimental results for response-time and throughput

As shown in Figure 5 the prediction performance ofSSRM outperforms other two approaches The performanceof IBCF is similar to UBCF As the values of 119896 increase betteraccuracy can be achieved This indicates that more similarneighbor records can provide more information for missingvalue prediction However when 119896 is larger than 25 MAEand RSME fail to drop with the value of 119896 increasing Themain reason is the limited number of similar neighbors Theobservations also show that RMSE has the similar trend butwith larger fluctuations

Next we investigate the impact of matrix density Thematrix density varies from 10 to 50 in increment of10 Similar user size k is set to 5 in this experimentFigure 6 shows the experimental results for response-timeand throughput

As presented in Figure 6 the prediction performance isalso enhanced with the value of matrix density increasingThe main reason is that denser user-service matrix providesmore information for missing value prediction NARM hasthe better performance than LBCF and UBCF under allexperimental setting consistently since NARM considerscontext similarity and the relational degree among cloudservices The experimental results of LBCF and UBCF aresimilar in this experiment since these two similarity compu-tation methods are similar to each other and are both rating-oriented

43 Performance Comparison with Other Popular ApproachesTo study the prediction performance of SSRM we com-pare SSRM with three existing QoS properties predictionapproaches CloudRank2 [3] TECSS [4] and JV-PCC [5]

The size of Top-119896 similar service is an important factorin CF approach which determines how many neighborsrsquohistorical records are employed to generate predictions [5]Therefore the impact of matrix density is not considered in

the experiments We set the density to 10 We vary 119896 from 5to 30 in increment of 5 In order to simplify the experimentalprocess we consider all cloud services are trustworthyFigure 7 shows the experimental results for response-timeand throughput Under the same simulation condition SSRMsignificantly outperforms CloudRank2 TECSS and JV-PCCThe observations also suggest that better accuracy can beachieved by our model when more historical records areavailable in the service selection studyThemain reason is thatcomputing context similarity can improve the performance ofprediction evaluation in MCC environments

Figures 7(a) and 7(b) depict the MAE fractions of SSRMCloudRank2 TECSS and JV-PCC for response-time andthroughput while Figures 7(c) and 7(d) depict the RMSEfractions It can be observed that SSRM achieves smallerMAE and RMSE consistently than other approaches for bothresponse-time and throughput

5 Conclusions and Future

In the paper we propose a novel service selection and recom-mendation model (SSRM) for mobile cloud computing envi-ronmentsWhen calculating user similarity context informa-tion and interest are considered Through utilizing relationaldegree to improve the accuracy of ranking result our serviceselection and recommendation method succeed in obtainingthe ultimate service ranking results In addition SSRM onlyselects trusted mobile cloud service as candidates Thereforea set of trusted candidates are identified for target user Theexperimental results show that our approach significantlyimproves the prediction performance as comparedwith othertwo traditional approaches item-based CF approach (IBCF)and user-based CF approach (UBCF)

There are some disadvantages in our SSRM Firstly weexploit node distance to compute context similarity Thereis an underlying assumption in this exploitation each ofthe two adjacent nodes has equal semantic distance orgranularity of nodes in each level is identicalThis underlyingassumption is not true in some scenarios and it deterioratesthe performance of similarity evaluation Secondly the trustthreshold is used to select trusted mobile cloud service ascandidates However it cannot detect and exclude maliciousQoS values provided by users Thirdly in our experimentsthere are only three types of context information that areconsidered

Therefore in the future we will study the methods toimprove context similarity measurement We will study how

Wireless Communications and Mobile Computing 11

Response-time

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E10 25 3015 205

k

(b)Response-time

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

(c)

Throughput

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

(d)

Figure 5 Impact of neighbor size 119896 comparison with SSRM IBCF and UBCF

12 Wireless Communications and Mobile Computing

Response-time

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E

20 50 6030 4010Matrix density ()

(a)

Throughput

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E20 30 40 50 6010

Matrix density ()

(b)

Response-time

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 50 6030 4010Matrix density ()

(c)

Throughput

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 30 40 50 6010Matrix density ()

(d)

Figure 6 Impact of matrix density comparison with SSRM IBCF and UBCF

Wireless Communications and Mobile Computing 13

Response-time

JV-PCCTECSSSSRM

CloudRank2

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

02

025

03

035

04

045

05

055

06

065

07

MA

E10 15 20 25 305

k

JV-PCCTECSSSSRM

CloudRank2

(b)

Response-time

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

JV-PCCTECSSSSRM

CloudRank2

(c)

Throughput

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

JV-PCCTECSSSSRM

CloudRank2

(d)

Figure 7 Impact of neighbor size 119896 comparison with other popular approaches

14 Wireless Communications and Mobile Computing

to select themost trustworthy cloud service of certain type forthe active users based on multidimensional trust evidenceWe also will conduct more experimental investigations thatdeal with the impact of different context changes on serviceselection Moreover we will improve the ranking accuracyof our approaches by exploiting additional techniques (com-bining rating-oriented approaches and ranking-orientedapproaches matrix factorization intelligent optimizationalgorithms etc)

Conflicts of Interest

The author declares that they have no conflicts of interest

Acknowledgments

The work in this paper has been supported by NationalNatural Science Foundation of China (Program no 71501156)and China Postdoctoral Science Foundation (Program no2014M560796) and Shanxi Provincial EducationDepartment(Program no 15JK1679)

References

[1] White PaperMobile CloudComputing Solution Brief AEPONA2010

[2] httpsenwikipediaorgwikiCloudlet[3] Z Zheng X Wu Y Zhang M R Lyu and J Wang ldquoQoS

ranking prediction for cloud servicesrdquo IEEE Transactions onParallel and Distributed Systems vol 24 no 6 pp 1213ndash12222013

[4] Y Pan SDingW Fan J Li and S Yang ldquoTrust-enhanced cloudservice selection model based on QoS analysisrdquo PLoS ONE vol10 no 11 Article ID e0143448 2015

[5] S Ding C-Y Xia K-L Zhou S-L Yang and J S Shang ldquoDeci-sion support for personalized cloud service selection throughmulti-attribute trustworthiness evaluationrdquo PLoS ONE vol 9no 6 Article ID e107821 2014

[6] S K Garg S Versteeg and R Buyya ldquoA framework for rankingof cloud computing servicesrdquo Future Generation ComputerSystems vol 29 no 4 pp 1012ndash1023 2013

[7] H Ma Z Hu K Li and H Zhang ldquoToward trustworthy cloudservice selection a time-aware approach using interval neutro-sophic setrdquo Journal of Parallel and Distributed Computing vol96 pp 75ndash94 2016

[8] A Ahmed and A S Sabyasachi ldquoMobile cloud service selectionusing back propagation neural networkrdquo International Journalof Computer Applications vol 129 no 14 pp 1ndash5 2015

[9] L Guo J Ma Z-M Chen and H-R Jiang ldquoIncorporatingitem relations for social recommendationrdquo Chinese Journal ofComputers vol 37 no 1 pp 219ndash228 2014

[10] Z Yang B Wu K Zheng X Wang and L Lei ldquoA survey ofcollaborative filtering-based recommender systems for mobileinternet applicationsrdquo IEEE Access vol 4 pp 3273ndash3287 2016

[11] Markets and Markets httpwwwmarketsandmarketscomPressReleasesmobile-cloudasp

[12] M Whaiduzzaman A Gani N B Anuar M Shiraz MN Haque and I T Haque ldquoCloud service selection usingmulticriteria decision analysisrdquoTheScientificWorld Journal vol2014 Article ID 459375 10 pages 2014

[13] K Tserpes F Aisopos D Kyriazis and T Varvarigou ldquoArecommender mechanism for service selection in service-oriented environmentsrdquo Future Generation Computer Systemsvol 28 no 8 pp 1285ndash1294 2012

[14] Z U Rehman O K Hussain and F K Hussain ldquoParallelcloud service selection and ranking based on QoS historyrdquoInternational Journal of Parallel Programming vol 42 no 5 pp820ndash852 2014

[15] W-J Fan S-L YangH Perros and J Pei ldquoAmulti-dimensionaltrust-aware cloud service selection mechanism based on evi-dential reasoning approachrdquo International Journal of Automa-tion and Computing vol 12 no 2 pp 208ndash219 2015

[16] X G Wang J Cao and Y Xiang ldquoDynamic cloud serviceselection using an adaptive learning mechanism in multi-cloudcomputingrdquo The Journal of Systems and Software vol 100 pp195ndash210 2015

[17] C T Do N H Tran E-N Huh C S Hong D Niyato andZ Han ldquoDynamics of service selection and provider pricinggame in heterogeneous cloud marketrdquo Journal of Network andComputer Applications vol 69 pp 152ndash165 2015

[18] R Pal and P Hui ldquoEconomic models for cloud service marketspricing and capacity planningrdquo Theoretical Computer Sciencevol 496 pp 113ndash124 2013

[19] A Ezenwoke O Daramola and M Adigun ldquoTowards avisualization framework for service selection in cloud e-marketplacesrdquo in Proceedings of the 2017 IEEE InternationalConference on Web Services (ICWS rsquo17) pp 828ndash835 HonoluluHI USA June 2017

[20] M Tang X Dai J Liu and J Chen ldquoTowards a trust evaluationmiddleware for cloud service selectionrdquo Future GenerationComputer Systems vol 74 pp 302ndash312 2017

[21] A K Dey ldquoUnderstanding and using contextrdquo Personal andUbiquitous Computing vol 5 no 1 pp 4ndash7 2001

[22] X Han L Wang S Park A Cuevas and N Crespi ldquoAlikepeople alike interests A large-scale study on interest similarityin social networksrdquo in Proceedings of the 2014 IEEEACM Inter-national Conference on Advances in Social Networks Analysisand Mining (ASONAM rsquo14) pp 491ndash496 August 2014

[23] R N Lichtenwalter J T Lussier and N V Chawla ldquoNewperspectives and methods in link predictionrdquo in Proceedings ofthe 16th ACM SIGKDD International Conference on KnowledgeDiscovery and Data Mining (KDD rsquo10) pp 243ndash252 July 2010

[24] L Munasinghe and R Ichise ldquoLink prediction in social net-works using information flow via active linksrdquo IEICE Transac-tion on Information and Systems vol E96-D no 7 pp 1495ndash1502 2013

[25] Axis httpaxisapacheorgaxis2[26] ns3 httpswwwnsnamorgns3[27] Dataset httpwwwzibinzhengcomtpds

Wireless Communications and Mobile Computing 9

Response-time

02

025

03

035

04

045

05

055

06

065

07

MA

E

01 02 03 04 05 06 07 08 09 100Weighting factor

(a) MAE values

Response-time

03

035

04

045

05

055

06

065

07

075

08

RMSE

01 02 03 04 05 06 07 08 09 100Weighting factor

(b) RMSE values

Figure 4 Impact of weighting factor 120582

Table 1 Mobile cloud service QoS dataset and context informationdescriptions

Statistics ValueNumber of mobile cloud serviceinvocations 180000

Number of service users 100Number of mobile cloud services 200Minimum response-time value 0004 sMaximum response-time value 20 sMean of response-time 110 sStandard deviation ofresponse-time 226 s

Minimum throughput value 01 kbpsMaximum throughput value 1000 kbpsMean of throughput 2546 kbpsStandard deviation of throughput 4803 kbpsBandwidth 4Mndash8MMemory 100000000MBndash2100000000MBBattery capacity 05 KVAndash2KVA

where 119903act119906119909mcs119895 and 119903pre119906119909mcs119895 denote the actual QoS valueand predicted value respectively 119897pre is the number ofpredicted QoS values Smaller values of MAE and RMSEindicate better results

41 Impact of 120582 120582 is weighting factor in (2) 120582 determineshow much similarity measure relies on context and interestIn the experiment we vary the weighting factor 120582 from 0 to 1in increment of 01 The Top-k is set to 5 Matrix density is setto 10 Matrix density means the percentage of selected QoSentries that are used to predict the missing QoS value Theexperimental results are shown in Figure 4 As the value of 120582increases MAE firstly decreases and then quickly increasesWhen 120582 = 04 our results have the best value As shown inFigure 4(b) RMSE presents similar trend

42 Performance Comparison of SSRM IBCF and UBCFTo compare the performance of SSRM IBCF and UBCFsimilarity computation methods we implement IBCF andUBCF methods

Item-Based CF Approach (IBCF) We apply PCC to calculatesimilarities between users and predicted QoS values based onsimilar users The user similarity is calculated by

sim (119906119909 119906119910) = summcs119895isinmcs119906119909119906119910(119903119906119909 mcs119895 minus 119903119906119909) (119903119906119910 mcs119895 minus 119903119906119910)

radicsummcs119895isinmcs119906119909119906119910(119903119906119909mcs119895 minus 119903119906119909)2radicsummcs119895isinmcs119906119909119906119910

(119903119906119910mcs119895 minus 119903119906119910)2 (18)

10 Wireless Communications and Mobile Computing

where mcs119906119909119906119910 is the set of mobile cloud services that havebeen coinvoked by 119906119909 and 119906119910 119903119906119909and 119903119906119910 are average QoSvalues of cloud service invoked by 119906119909 and 119906119910 respectively

User-Based CF Approach (UBCF) We apply PCC to calculatesimilarities between services and predicted QoS values basedon similar services The service similarity is calculated by

sim (mcs119895mcs119894) = sum119906119909isin119906mcs119895mcs119894(119903119906119909mcs119895 minus 119903mcs119895) (119903119906119909 mcs119894 minus 119903mcs119894)

radicsum119906119909isin119906mcs119895mcs119894(119903119906119909mcs119895 minus 119903mcs119895)2radicsum119906119909isin119906mcs119895mcs119894

(119903119906119909mcs119894 minus 119903mcs119894)2 (19)

where 119906mcs119895mcs119894 is the set of users who have invoked the cloudservices mcs119895 andmcs119894 119903mcs119895 and 119903mcs119894 are average QoS valuesof cloud services mcs119895 and mcs119894 made by 119906119909

In the experiments trustworthiness of cloud service hasno influence on similarity computation and service decisionin the experiments In order to simplify the experimentalprocess we consider all cloud services are trustworthyWeighting factor 120582 is set to 04 We firstly study the effectof neighbor size 119896 Matrix density is set to 10 We change119896 from 5 to 30 in increment of 5 Figure 5 shows theexperimental results for response-time and throughput

As shown in Figure 5 the prediction performance ofSSRM outperforms other two approaches The performanceof IBCF is similar to UBCF As the values of 119896 increase betteraccuracy can be achieved This indicates that more similarneighbor records can provide more information for missingvalue prediction However when 119896 is larger than 25 MAEand RSME fail to drop with the value of 119896 increasing Themain reason is the limited number of similar neighbors Theobservations also show that RMSE has the similar trend butwith larger fluctuations

Next we investigate the impact of matrix density Thematrix density varies from 10 to 50 in increment of10 Similar user size k is set to 5 in this experimentFigure 6 shows the experimental results for response-timeand throughput

As presented in Figure 6 the prediction performance isalso enhanced with the value of matrix density increasingThe main reason is that denser user-service matrix providesmore information for missing value prediction NARM hasthe better performance than LBCF and UBCF under allexperimental setting consistently since NARM considerscontext similarity and the relational degree among cloudservices The experimental results of LBCF and UBCF aresimilar in this experiment since these two similarity compu-tation methods are similar to each other and are both rating-oriented

43 Performance Comparison with Other Popular ApproachesTo study the prediction performance of SSRM we com-pare SSRM with three existing QoS properties predictionapproaches CloudRank2 [3] TECSS [4] and JV-PCC [5]

The size of Top-119896 similar service is an important factorin CF approach which determines how many neighborsrsquohistorical records are employed to generate predictions [5]Therefore the impact of matrix density is not considered in

the experiments We set the density to 10 We vary 119896 from 5to 30 in increment of 5 In order to simplify the experimentalprocess we consider all cloud services are trustworthyFigure 7 shows the experimental results for response-timeand throughput Under the same simulation condition SSRMsignificantly outperforms CloudRank2 TECSS and JV-PCCThe observations also suggest that better accuracy can beachieved by our model when more historical records areavailable in the service selection studyThemain reason is thatcomputing context similarity can improve the performance ofprediction evaluation in MCC environments

Figures 7(a) and 7(b) depict the MAE fractions of SSRMCloudRank2 TECSS and JV-PCC for response-time andthroughput while Figures 7(c) and 7(d) depict the RMSEfractions It can be observed that SSRM achieves smallerMAE and RMSE consistently than other approaches for bothresponse-time and throughput

5 Conclusions and Future

In the paper we propose a novel service selection and recom-mendation model (SSRM) for mobile cloud computing envi-ronmentsWhen calculating user similarity context informa-tion and interest are considered Through utilizing relationaldegree to improve the accuracy of ranking result our serviceselection and recommendation method succeed in obtainingthe ultimate service ranking results In addition SSRM onlyselects trusted mobile cloud service as candidates Thereforea set of trusted candidates are identified for target user Theexperimental results show that our approach significantlyimproves the prediction performance as comparedwith othertwo traditional approaches item-based CF approach (IBCF)and user-based CF approach (UBCF)

There are some disadvantages in our SSRM Firstly weexploit node distance to compute context similarity Thereis an underlying assumption in this exploitation each ofthe two adjacent nodes has equal semantic distance orgranularity of nodes in each level is identicalThis underlyingassumption is not true in some scenarios and it deterioratesthe performance of similarity evaluation Secondly the trustthreshold is used to select trusted mobile cloud service ascandidates However it cannot detect and exclude maliciousQoS values provided by users Thirdly in our experimentsthere are only three types of context information that areconsidered

Therefore in the future we will study the methods toimprove context similarity measurement We will study how

Wireless Communications and Mobile Computing 11

Response-time

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E10 25 3015 205

k

(b)Response-time

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

(c)

Throughput

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

(d)

Figure 5 Impact of neighbor size 119896 comparison with SSRM IBCF and UBCF

12 Wireless Communications and Mobile Computing

Response-time

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E

20 50 6030 4010Matrix density ()

(a)

Throughput

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E20 30 40 50 6010

Matrix density ()

(b)

Response-time

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 50 6030 4010Matrix density ()

(c)

Throughput

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 30 40 50 6010Matrix density ()

(d)

Figure 6 Impact of matrix density comparison with SSRM IBCF and UBCF

Wireless Communications and Mobile Computing 13

Response-time

JV-PCCTECSSSSRM

CloudRank2

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

02

025

03

035

04

045

05

055

06

065

07

MA

E10 15 20 25 305

k

JV-PCCTECSSSSRM

CloudRank2

(b)

Response-time

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

JV-PCCTECSSSSRM

CloudRank2

(c)

Throughput

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

JV-PCCTECSSSSRM

CloudRank2

(d)

Figure 7 Impact of neighbor size 119896 comparison with other popular approaches

14 Wireless Communications and Mobile Computing

to select themost trustworthy cloud service of certain type forthe active users based on multidimensional trust evidenceWe also will conduct more experimental investigations thatdeal with the impact of different context changes on serviceselection Moreover we will improve the ranking accuracyof our approaches by exploiting additional techniques (com-bining rating-oriented approaches and ranking-orientedapproaches matrix factorization intelligent optimizationalgorithms etc)

Conflicts of Interest

The author declares that they have no conflicts of interest

Acknowledgments

The work in this paper has been supported by NationalNatural Science Foundation of China (Program no 71501156)and China Postdoctoral Science Foundation (Program no2014M560796) and Shanxi Provincial EducationDepartment(Program no 15JK1679)

References

[1] White PaperMobile CloudComputing Solution Brief AEPONA2010

[2] httpsenwikipediaorgwikiCloudlet[3] Z Zheng X Wu Y Zhang M R Lyu and J Wang ldquoQoS

ranking prediction for cloud servicesrdquo IEEE Transactions onParallel and Distributed Systems vol 24 no 6 pp 1213ndash12222013

[4] Y Pan SDingW Fan J Li and S Yang ldquoTrust-enhanced cloudservice selection model based on QoS analysisrdquo PLoS ONE vol10 no 11 Article ID e0143448 2015

[5] S Ding C-Y Xia K-L Zhou S-L Yang and J S Shang ldquoDeci-sion support for personalized cloud service selection throughmulti-attribute trustworthiness evaluationrdquo PLoS ONE vol 9no 6 Article ID e107821 2014

[6] S K Garg S Versteeg and R Buyya ldquoA framework for rankingof cloud computing servicesrdquo Future Generation ComputerSystems vol 29 no 4 pp 1012ndash1023 2013

[7] H Ma Z Hu K Li and H Zhang ldquoToward trustworthy cloudservice selection a time-aware approach using interval neutro-sophic setrdquo Journal of Parallel and Distributed Computing vol96 pp 75ndash94 2016

[8] A Ahmed and A S Sabyasachi ldquoMobile cloud service selectionusing back propagation neural networkrdquo International Journalof Computer Applications vol 129 no 14 pp 1ndash5 2015

[9] L Guo J Ma Z-M Chen and H-R Jiang ldquoIncorporatingitem relations for social recommendationrdquo Chinese Journal ofComputers vol 37 no 1 pp 219ndash228 2014

[10] Z Yang B Wu K Zheng X Wang and L Lei ldquoA survey ofcollaborative filtering-based recommender systems for mobileinternet applicationsrdquo IEEE Access vol 4 pp 3273ndash3287 2016

[11] Markets and Markets httpwwwmarketsandmarketscomPressReleasesmobile-cloudasp

[12] M Whaiduzzaman A Gani N B Anuar M Shiraz MN Haque and I T Haque ldquoCloud service selection usingmulticriteria decision analysisrdquoTheScientificWorld Journal vol2014 Article ID 459375 10 pages 2014

[13] K Tserpes F Aisopos D Kyriazis and T Varvarigou ldquoArecommender mechanism for service selection in service-oriented environmentsrdquo Future Generation Computer Systemsvol 28 no 8 pp 1285ndash1294 2012

[14] Z U Rehman O K Hussain and F K Hussain ldquoParallelcloud service selection and ranking based on QoS historyrdquoInternational Journal of Parallel Programming vol 42 no 5 pp820ndash852 2014

[15] W-J Fan S-L YangH Perros and J Pei ldquoAmulti-dimensionaltrust-aware cloud service selection mechanism based on evi-dential reasoning approachrdquo International Journal of Automa-tion and Computing vol 12 no 2 pp 208ndash219 2015

[16] X G Wang J Cao and Y Xiang ldquoDynamic cloud serviceselection using an adaptive learning mechanism in multi-cloudcomputingrdquo The Journal of Systems and Software vol 100 pp195ndash210 2015

[17] C T Do N H Tran E-N Huh C S Hong D Niyato andZ Han ldquoDynamics of service selection and provider pricinggame in heterogeneous cloud marketrdquo Journal of Network andComputer Applications vol 69 pp 152ndash165 2015

[18] R Pal and P Hui ldquoEconomic models for cloud service marketspricing and capacity planningrdquo Theoretical Computer Sciencevol 496 pp 113ndash124 2013

[19] A Ezenwoke O Daramola and M Adigun ldquoTowards avisualization framework for service selection in cloud e-marketplacesrdquo in Proceedings of the 2017 IEEE InternationalConference on Web Services (ICWS rsquo17) pp 828ndash835 HonoluluHI USA June 2017

[20] M Tang X Dai J Liu and J Chen ldquoTowards a trust evaluationmiddleware for cloud service selectionrdquo Future GenerationComputer Systems vol 74 pp 302ndash312 2017

[21] A K Dey ldquoUnderstanding and using contextrdquo Personal andUbiquitous Computing vol 5 no 1 pp 4ndash7 2001

[22] X Han L Wang S Park A Cuevas and N Crespi ldquoAlikepeople alike interests A large-scale study on interest similarityin social networksrdquo in Proceedings of the 2014 IEEEACM Inter-national Conference on Advances in Social Networks Analysisand Mining (ASONAM rsquo14) pp 491ndash496 August 2014

[23] R N Lichtenwalter J T Lussier and N V Chawla ldquoNewperspectives and methods in link predictionrdquo in Proceedings ofthe 16th ACM SIGKDD International Conference on KnowledgeDiscovery and Data Mining (KDD rsquo10) pp 243ndash252 July 2010

[24] L Munasinghe and R Ichise ldquoLink prediction in social net-works using information flow via active linksrdquo IEICE Transac-tion on Information and Systems vol E96-D no 7 pp 1495ndash1502 2013

[25] Axis httpaxisapacheorgaxis2[26] ns3 httpswwwnsnamorgns3[27] Dataset httpwwwzibinzhengcomtpds

10 Wireless Communications and Mobile Computing

where mcs119906119909119906119910 is the set of mobile cloud services that havebeen coinvoked by 119906119909 and 119906119910 119903119906119909and 119903119906119910 are average QoSvalues of cloud service invoked by 119906119909 and 119906119910 respectively

User-Based CF Approach (UBCF) We apply PCC to calculatesimilarities between services and predicted QoS values basedon similar services The service similarity is calculated by

sim (mcs119895mcs119894) = sum119906119909isin119906mcs119895mcs119894(119903119906119909mcs119895 minus 119903mcs119895) (119903119906119909 mcs119894 minus 119903mcs119894)

radicsum119906119909isin119906mcs119895mcs119894(119903119906119909mcs119895 minus 119903mcs119895)2radicsum119906119909isin119906mcs119895mcs119894

(119903119906119909mcs119894 minus 119903mcs119894)2 (19)

where 119906mcs119895mcs119894 is the set of users who have invoked the cloudservices mcs119895 andmcs119894 119903mcs119895 and 119903mcs119894 are average QoS valuesof cloud services mcs119895 and mcs119894 made by 119906119909

In the experiments trustworthiness of cloud service hasno influence on similarity computation and service decisionin the experiments In order to simplify the experimentalprocess we consider all cloud services are trustworthyWeighting factor 120582 is set to 04 We firstly study the effectof neighbor size 119896 Matrix density is set to 10 We change119896 from 5 to 30 in increment of 5 Figure 5 shows theexperimental results for response-time and throughput

As shown in Figure 5 the prediction performance ofSSRM outperforms other two approaches The performanceof IBCF is similar to UBCF As the values of 119896 increase betteraccuracy can be achieved This indicates that more similarneighbor records can provide more information for missingvalue prediction However when 119896 is larger than 25 MAEand RSME fail to drop with the value of 119896 increasing Themain reason is the limited number of similar neighbors Theobservations also show that RMSE has the similar trend butwith larger fluctuations

Next we investigate the impact of matrix density Thematrix density varies from 10 to 50 in increment of10 Similar user size k is set to 5 in this experimentFigure 6 shows the experimental results for response-timeand throughput

As presented in Figure 6 the prediction performance isalso enhanced with the value of matrix density increasingThe main reason is that denser user-service matrix providesmore information for missing value prediction NARM hasthe better performance than LBCF and UBCF under allexperimental setting consistently since NARM considerscontext similarity and the relational degree among cloudservices The experimental results of LBCF and UBCF aresimilar in this experiment since these two similarity compu-tation methods are similar to each other and are both rating-oriented

43 Performance Comparison with Other Popular ApproachesTo study the prediction performance of SSRM we com-pare SSRM with three existing QoS properties predictionapproaches CloudRank2 [3] TECSS [4] and JV-PCC [5]

The size of Top-119896 similar service is an important factorin CF approach which determines how many neighborsrsquohistorical records are employed to generate predictions [5]Therefore the impact of matrix density is not considered in

the experiments We set the density to 10 We vary 119896 from 5to 30 in increment of 5 In order to simplify the experimentalprocess we consider all cloud services are trustworthyFigure 7 shows the experimental results for response-timeand throughput Under the same simulation condition SSRMsignificantly outperforms CloudRank2 TECSS and JV-PCCThe observations also suggest that better accuracy can beachieved by our model when more historical records areavailable in the service selection studyThemain reason is thatcomputing context similarity can improve the performance ofprediction evaluation in MCC environments

Figures 7(a) and 7(b) depict the MAE fractions of SSRMCloudRank2 TECSS and JV-PCC for response-time andthroughput while Figures 7(c) and 7(d) depict the RMSEfractions It can be observed that SSRM achieves smallerMAE and RMSE consistently than other approaches for bothresponse-time and throughput

5 Conclusions and Future

In the paper we propose a novel service selection and recom-mendation model (SSRM) for mobile cloud computing envi-ronmentsWhen calculating user similarity context informa-tion and interest are considered Through utilizing relationaldegree to improve the accuracy of ranking result our serviceselection and recommendation method succeed in obtainingthe ultimate service ranking results In addition SSRM onlyselects trusted mobile cloud service as candidates Thereforea set of trusted candidates are identified for target user Theexperimental results show that our approach significantlyimproves the prediction performance as comparedwith othertwo traditional approaches item-based CF approach (IBCF)and user-based CF approach (UBCF)

There are some disadvantages in our SSRM Firstly weexploit node distance to compute context similarity Thereis an underlying assumption in this exploitation each ofthe two adjacent nodes has equal semantic distance orgranularity of nodes in each level is identicalThis underlyingassumption is not true in some scenarios and it deterioratesthe performance of similarity evaluation Secondly the trustthreshold is used to select trusted mobile cloud service ascandidates However it cannot detect and exclude maliciousQoS values provided by users Thirdly in our experimentsthere are only three types of context information that areconsidered

Therefore in the future we will study the methods toimprove context similarity measurement We will study how

Wireless Communications and Mobile Computing 11

Response-time

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E10 25 3015 205

k

(b)Response-time

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

(c)

Throughput

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

(d)

Figure 5 Impact of neighbor size 119896 comparison with SSRM IBCF and UBCF

12 Wireless Communications and Mobile Computing

Response-time

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E

20 50 6030 4010Matrix density ()

(a)

Throughput

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E20 30 40 50 6010

Matrix density ()

(b)

Response-time

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 50 6030 4010Matrix density ()

(c)

Throughput

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 30 40 50 6010Matrix density ()

(d)

Figure 6 Impact of matrix density comparison with SSRM IBCF and UBCF

Wireless Communications and Mobile Computing 13

Response-time

JV-PCCTECSSSSRM

CloudRank2

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

02

025

03

035

04

045

05

055

06

065

07

MA

E10 15 20 25 305

k

JV-PCCTECSSSSRM

CloudRank2

(b)

Response-time

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

JV-PCCTECSSSSRM

CloudRank2

(c)

Throughput

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

JV-PCCTECSSSSRM

CloudRank2

(d)

Figure 7 Impact of neighbor size 119896 comparison with other popular approaches

14 Wireless Communications and Mobile Computing

to select themost trustworthy cloud service of certain type forthe active users based on multidimensional trust evidenceWe also will conduct more experimental investigations thatdeal with the impact of different context changes on serviceselection Moreover we will improve the ranking accuracyof our approaches by exploiting additional techniques (com-bining rating-oriented approaches and ranking-orientedapproaches matrix factorization intelligent optimizationalgorithms etc)

Conflicts of Interest

The author declares that they have no conflicts of interest

Acknowledgments

The work in this paper has been supported by NationalNatural Science Foundation of China (Program no 71501156)and China Postdoctoral Science Foundation (Program no2014M560796) and Shanxi Provincial EducationDepartment(Program no 15JK1679)

References

[1] White PaperMobile CloudComputing Solution Brief AEPONA2010

[2] httpsenwikipediaorgwikiCloudlet[3] Z Zheng X Wu Y Zhang M R Lyu and J Wang ldquoQoS

ranking prediction for cloud servicesrdquo IEEE Transactions onParallel and Distributed Systems vol 24 no 6 pp 1213ndash12222013

[4] Y Pan SDingW Fan J Li and S Yang ldquoTrust-enhanced cloudservice selection model based on QoS analysisrdquo PLoS ONE vol10 no 11 Article ID e0143448 2015

[5] S Ding C-Y Xia K-L Zhou S-L Yang and J S Shang ldquoDeci-sion support for personalized cloud service selection throughmulti-attribute trustworthiness evaluationrdquo PLoS ONE vol 9no 6 Article ID e107821 2014

[6] S K Garg S Versteeg and R Buyya ldquoA framework for rankingof cloud computing servicesrdquo Future Generation ComputerSystems vol 29 no 4 pp 1012ndash1023 2013

[7] H Ma Z Hu K Li and H Zhang ldquoToward trustworthy cloudservice selection a time-aware approach using interval neutro-sophic setrdquo Journal of Parallel and Distributed Computing vol96 pp 75ndash94 2016

[8] A Ahmed and A S Sabyasachi ldquoMobile cloud service selectionusing back propagation neural networkrdquo International Journalof Computer Applications vol 129 no 14 pp 1ndash5 2015

[9] L Guo J Ma Z-M Chen and H-R Jiang ldquoIncorporatingitem relations for social recommendationrdquo Chinese Journal ofComputers vol 37 no 1 pp 219ndash228 2014

[10] Z Yang B Wu K Zheng X Wang and L Lei ldquoA survey ofcollaborative filtering-based recommender systems for mobileinternet applicationsrdquo IEEE Access vol 4 pp 3273ndash3287 2016

[11] Markets and Markets httpwwwmarketsandmarketscomPressReleasesmobile-cloudasp

[12] M Whaiduzzaman A Gani N B Anuar M Shiraz MN Haque and I T Haque ldquoCloud service selection usingmulticriteria decision analysisrdquoTheScientificWorld Journal vol2014 Article ID 459375 10 pages 2014

[13] K Tserpes F Aisopos D Kyriazis and T Varvarigou ldquoArecommender mechanism for service selection in service-oriented environmentsrdquo Future Generation Computer Systemsvol 28 no 8 pp 1285ndash1294 2012

[14] Z U Rehman O K Hussain and F K Hussain ldquoParallelcloud service selection and ranking based on QoS historyrdquoInternational Journal of Parallel Programming vol 42 no 5 pp820ndash852 2014

[15] W-J Fan S-L YangH Perros and J Pei ldquoAmulti-dimensionaltrust-aware cloud service selection mechanism based on evi-dential reasoning approachrdquo International Journal of Automa-tion and Computing vol 12 no 2 pp 208ndash219 2015

[16] X G Wang J Cao and Y Xiang ldquoDynamic cloud serviceselection using an adaptive learning mechanism in multi-cloudcomputingrdquo The Journal of Systems and Software vol 100 pp195ndash210 2015

[17] C T Do N H Tran E-N Huh C S Hong D Niyato andZ Han ldquoDynamics of service selection and provider pricinggame in heterogeneous cloud marketrdquo Journal of Network andComputer Applications vol 69 pp 152ndash165 2015

[18] R Pal and P Hui ldquoEconomic models for cloud service marketspricing and capacity planningrdquo Theoretical Computer Sciencevol 496 pp 113ndash124 2013

[19] A Ezenwoke O Daramola and M Adigun ldquoTowards avisualization framework for service selection in cloud e-marketplacesrdquo in Proceedings of the 2017 IEEE InternationalConference on Web Services (ICWS rsquo17) pp 828ndash835 HonoluluHI USA June 2017

[20] M Tang X Dai J Liu and J Chen ldquoTowards a trust evaluationmiddleware for cloud service selectionrdquo Future GenerationComputer Systems vol 74 pp 302ndash312 2017

[21] A K Dey ldquoUnderstanding and using contextrdquo Personal andUbiquitous Computing vol 5 no 1 pp 4ndash7 2001

[22] X Han L Wang S Park A Cuevas and N Crespi ldquoAlikepeople alike interests A large-scale study on interest similarityin social networksrdquo in Proceedings of the 2014 IEEEACM Inter-national Conference on Advances in Social Networks Analysisand Mining (ASONAM rsquo14) pp 491ndash496 August 2014

[23] R N Lichtenwalter J T Lussier and N V Chawla ldquoNewperspectives and methods in link predictionrdquo in Proceedings ofthe 16th ACM SIGKDD International Conference on KnowledgeDiscovery and Data Mining (KDD rsquo10) pp 243ndash252 July 2010

[24] L Munasinghe and R Ichise ldquoLink prediction in social net-works using information flow via active linksrdquo IEICE Transac-tion on Information and Systems vol E96-D no 7 pp 1495ndash1502 2013

[25] Axis httpaxisapacheorgaxis2[26] ns3 httpswwwnsnamorgns3[27] Dataset httpwwwzibinzhengcomtpds

Wireless Communications and Mobile Computing 11

Response-time

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E10 25 3015 205

k

(b)Response-time

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

(c)

Throughput

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

(d)

Figure 5 Impact of neighbor size 119896 comparison with SSRM IBCF and UBCF

12 Wireless Communications and Mobile Computing

Response-time

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E

20 50 6030 4010Matrix density ()

(a)

Throughput

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E20 30 40 50 6010

Matrix density ()

(b)

Response-time

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 50 6030 4010Matrix density ()

(c)

Throughput

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 30 40 50 6010Matrix density ()

(d)

Figure 6 Impact of matrix density comparison with SSRM IBCF and UBCF

Wireless Communications and Mobile Computing 13

Response-time

JV-PCCTECSSSSRM

CloudRank2

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

02

025

03

035

04

045

05

055

06

065

07

MA

E10 15 20 25 305

k

JV-PCCTECSSSSRM

CloudRank2

(b)

Response-time

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

JV-PCCTECSSSSRM

CloudRank2

(c)

Throughput

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

JV-PCCTECSSSSRM

CloudRank2

(d)

Figure 7 Impact of neighbor size 119896 comparison with other popular approaches

14 Wireless Communications and Mobile Computing

to select themost trustworthy cloud service of certain type forthe active users based on multidimensional trust evidenceWe also will conduct more experimental investigations thatdeal with the impact of different context changes on serviceselection Moreover we will improve the ranking accuracyof our approaches by exploiting additional techniques (com-bining rating-oriented approaches and ranking-orientedapproaches matrix factorization intelligent optimizationalgorithms etc)

Conflicts of Interest

The author declares that they have no conflicts of interest

Acknowledgments

The work in this paper has been supported by NationalNatural Science Foundation of China (Program no 71501156)and China Postdoctoral Science Foundation (Program no2014M560796) and Shanxi Provincial EducationDepartment(Program no 15JK1679)

References

[1] White PaperMobile CloudComputing Solution Brief AEPONA2010

[2] httpsenwikipediaorgwikiCloudlet[3] Z Zheng X Wu Y Zhang M R Lyu and J Wang ldquoQoS

ranking prediction for cloud servicesrdquo IEEE Transactions onParallel and Distributed Systems vol 24 no 6 pp 1213ndash12222013

[4] Y Pan SDingW Fan J Li and S Yang ldquoTrust-enhanced cloudservice selection model based on QoS analysisrdquo PLoS ONE vol10 no 11 Article ID e0143448 2015

[5] S Ding C-Y Xia K-L Zhou S-L Yang and J S Shang ldquoDeci-sion support for personalized cloud service selection throughmulti-attribute trustworthiness evaluationrdquo PLoS ONE vol 9no 6 Article ID e107821 2014

[6] S K Garg S Versteeg and R Buyya ldquoA framework for rankingof cloud computing servicesrdquo Future Generation ComputerSystems vol 29 no 4 pp 1012ndash1023 2013

[7] H Ma Z Hu K Li and H Zhang ldquoToward trustworthy cloudservice selection a time-aware approach using interval neutro-sophic setrdquo Journal of Parallel and Distributed Computing vol96 pp 75ndash94 2016

[8] A Ahmed and A S Sabyasachi ldquoMobile cloud service selectionusing back propagation neural networkrdquo International Journalof Computer Applications vol 129 no 14 pp 1ndash5 2015

[9] L Guo J Ma Z-M Chen and H-R Jiang ldquoIncorporatingitem relations for social recommendationrdquo Chinese Journal ofComputers vol 37 no 1 pp 219ndash228 2014

[10] Z Yang B Wu K Zheng X Wang and L Lei ldquoA survey ofcollaborative filtering-based recommender systems for mobileinternet applicationsrdquo IEEE Access vol 4 pp 3273ndash3287 2016

[11] Markets and Markets httpwwwmarketsandmarketscomPressReleasesmobile-cloudasp

[12] M Whaiduzzaman A Gani N B Anuar M Shiraz MN Haque and I T Haque ldquoCloud service selection usingmulticriteria decision analysisrdquoTheScientificWorld Journal vol2014 Article ID 459375 10 pages 2014

[13] K Tserpes F Aisopos D Kyriazis and T Varvarigou ldquoArecommender mechanism for service selection in service-oriented environmentsrdquo Future Generation Computer Systemsvol 28 no 8 pp 1285ndash1294 2012

[14] Z U Rehman O K Hussain and F K Hussain ldquoParallelcloud service selection and ranking based on QoS historyrdquoInternational Journal of Parallel Programming vol 42 no 5 pp820ndash852 2014

[15] W-J Fan S-L YangH Perros and J Pei ldquoAmulti-dimensionaltrust-aware cloud service selection mechanism based on evi-dential reasoning approachrdquo International Journal of Automa-tion and Computing vol 12 no 2 pp 208ndash219 2015

[16] X G Wang J Cao and Y Xiang ldquoDynamic cloud serviceselection using an adaptive learning mechanism in multi-cloudcomputingrdquo The Journal of Systems and Software vol 100 pp195ndash210 2015

[17] C T Do N H Tran E-N Huh C S Hong D Niyato andZ Han ldquoDynamics of service selection and provider pricinggame in heterogeneous cloud marketrdquo Journal of Network andComputer Applications vol 69 pp 152ndash165 2015

[18] R Pal and P Hui ldquoEconomic models for cloud service marketspricing and capacity planningrdquo Theoretical Computer Sciencevol 496 pp 113ndash124 2013

[19] A Ezenwoke O Daramola and M Adigun ldquoTowards avisualization framework for service selection in cloud e-marketplacesrdquo in Proceedings of the 2017 IEEE InternationalConference on Web Services (ICWS rsquo17) pp 828ndash835 HonoluluHI USA June 2017

[20] M Tang X Dai J Liu and J Chen ldquoTowards a trust evaluationmiddleware for cloud service selectionrdquo Future GenerationComputer Systems vol 74 pp 302ndash312 2017

[21] A K Dey ldquoUnderstanding and using contextrdquo Personal andUbiquitous Computing vol 5 no 1 pp 4ndash7 2001

[22] X Han L Wang S Park A Cuevas and N Crespi ldquoAlikepeople alike interests A large-scale study on interest similarityin social networksrdquo in Proceedings of the 2014 IEEEACM Inter-national Conference on Advances in Social Networks Analysisand Mining (ASONAM rsquo14) pp 491ndash496 August 2014

[23] R N Lichtenwalter J T Lussier and N V Chawla ldquoNewperspectives and methods in link predictionrdquo in Proceedings ofthe 16th ACM SIGKDD International Conference on KnowledgeDiscovery and Data Mining (KDD rsquo10) pp 243ndash252 July 2010

[24] L Munasinghe and R Ichise ldquoLink prediction in social net-works using information flow via active linksrdquo IEICE Transac-tion on Information and Systems vol E96-D no 7 pp 1495ndash1502 2013

[25] Axis httpaxisapacheorgaxis2[26] ns3 httpswwwnsnamorgns3[27] Dataset httpwwwzibinzhengcomtpds

12 Wireless Communications and Mobile Computing

Response-time

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E

20 50 6030 4010Matrix density ()

(a)

Throughput

SSRMIBCFUBCF

02

025

03

035

04

045

05

055

06

065

07

MA

E20 30 40 50 6010

Matrix density ()

(b)

Response-time

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 50 6030 4010Matrix density ()

(c)

Throughput

SSRMIBCFUBCF

03

035

04

045

05

055

06

065

07

075

08

MA

E

20 30 40 50 6010Matrix density ()

(d)

Figure 6 Impact of matrix density comparison with SSRM IBCF and UBCF

Wireless Communications and Mobile Computing 13

Response-time

JV-PCCTECSSSSRM

CloudRank2

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

02

025

03

035

04

045

05

055

06

065

07

MA

E10 15 20 25 305

k

JV-PCCTECSSSSRM

CloudRank2

(b)

Response-time

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

JV-PCCTECSSSSRM

CloudRank2

(c)

Throughput

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

JV-PCCTECSSSSRM

CloudRank2

(d)

Figure 7 Impact of neighbor size 119896 comparison with other popular approaches

14 Wireless Communications and Mobile Computing

to select themost trustworthy cloud service of certain type forthe active users based on multidimensional trust evidenceWe also will conduct more experimental investigations thatdeal with the impact of different context changes on serviceselection Moreover we will improve the ranking accuracyof our approaches by exploiting additional techniques (com-bining rating-oriented approaches and ranking-orientedapproaches matrix factorization intelligent optimizationalgorithms etc)

Conflicts of Interest

The author declares that they have no conflicts of interest

Acknowledgments

The work in this paper has been supported by NationalNatural Science Foundation of China (Program no 71501156)and China Postdoctoral Science Foundation (Program no2014M560796) and Shanxi Provincial EducationDepartment(Program no 15JK1679)

References

[1] White PaperMobile CloudComputing Solution Brief AEPONA2010

[2] httpsenwikipediaorgwikiCloudlet[3] Z Zheng X Wu Y Zhang M R Lyu and J Wang ldquoQoS

ranking prediction for cloud servicesrdquo IEEE Transactions onParallel and Distributed Systems vol 24 no 6 pp 1213ndash12222013

[4] Y Pan SDingW Fan J Li and S Yang ldquoTrust-enhanced cloudservice selection model based on QoS analysisrdquo PLoS ONE vol10 no 11 Article ID e0143448 2015

[5] S Ding C-Y Xia K-L Zhou S-L Yang and J S Shang ldquoDeci-sion support for personalized cloud service selection throughmulti-attribute trustworthiness evaluationrdquo PLoS ONE vol 9no 6 Article ID e107821 2014

[6] S K Garg S Versteeg and R Buyya ldquoA framework for rankingof cloud computing servicesrdquo Future Generation ComputerSystems vol 29 no 4 pp 1012ndash1023 2013

[7] H Ma Z Hu K Li and H Zhang ldquoToward trustworthy cloudservice selection a time-aware approach using interval neutro-sophic setrdquo Journal of Parallel and Distributed Computing vol96 pp 75ndash94 2016

[8] A Ahmed and A S Sabyasachi ldquoMobile cloud service selectionusing back propagation neural networkrdquo International Journalof Computer Applications vol 129 no 14 pp 1ndash5 2015

[9] L Guo J Ma Z-M Chen and H-R Jiang ldquoIncorporatingitem relations for social recommendationrdquo Chinese Journal ofComputers vol 37 no 1 pp 219ndash228 2014

[10] Z Yang B Wu K Zheng X Wang and L Lei ldquoA survey ofcollaborative filtering-based recommender systems for mobileinternet applicationsrdquo IEEE Access vol 4 pp 3273ndash3287 2016

[11] Markets and Markets httpwwwmarketsandmarketscomPressReleasesmobile-cloudasp

[12] M Whaiduzzaman A Gani N B Anuar M Shiraz MN Haque and I T Haque ldquoCloud service selection usingmulticriteria decision analysisrdquoTheScientificWorld Journal vol2014 Article ID 459375 10 pages 2014

[13] K Tserpes F Aisopos D Kyriazis and T Varvarigou ldquoArecommender mechanism for service selection in service-oriented environmentsrdquo Future Generation Computer Systemsvol 28 no 8 pp 1285ndash1294 2012

[14] Z U Rehman O K Hussain and F K Hussain ldquoParallelcloud service selection and ranking based on QoS historyrdquoInternational Journal of Parallel Programming vol 42 no 5 pp820ndash852 2014

[15] W-J Fan S-L YangH Perros and J Pei ldquoAmulti-dimensionaltrust-aware cloud service selection mechanism based on evi-dential reasoning approachrdquo International Journal of Automa-tion and Computing vol 12 no 2 pp 208ndash219 2015

[16] X G Wang J Cao and Y Xiang ldquoDynamic cloud serviceselection using an adaptive learning mechanism in multi-cloudcomputingrdquo The Journal of Systems and Software vol 100 pp195ndash210 2015

[17] C T Do N H Tran E-N Huh C S Hong D Niyato andZ Han ldquoDynamics of service selection and provider pricinggame in heterogeneous cloud marketrdquo Journal of Network andComputer Applications vol 69 pp 152ndash165 2015

[18] R Pal and P Hui ldquoEconomic models for cloud service marketspricing and capacity planningrdquo Theoretical Computer Sciencevol 496 pp 113ndash124 2013

[19] A Ezenwoke O Daramola and M Adigun ldquoTowards avisualization framework for service selection in cloud e-marketplacesrdquo in Proceedings of the 2017 IEEE InternationalConference on Web Services (ICWS rsquo17) pp 828ndash835 HonoluluHI USA June 2017

[20] M Tang X Dai J Liu and J Chen ldquoTowards a trust evaluationmiddleware for cloud service selectionrdquo Future GenerationComputer Systems vol 74 pp 302ndash312 2017

[21] A K Dey ldquoUnderstanding and using contextrdquo Personal andUbiquitous Computing vol 5 no 1 pp 4ndash7 2001

[22] X Han L Wang S Park A Cuevas and N Crespi ldquoAlikepeople alike interests A large-scale study on interest similarityin social networksrdquo in Proceedings of the 2014 IEEEACM Inter-national Conference on Advances in Social Networks Analysisand Mining (ASONAM rsquo14) pp 491ndash496 August 2014

[23] R N Lichtenwalter J T Lussier and N V Chawla ldquoNewperspectives and methods in link predictionrdquo in Proceedings ofthe 16th ACM SIGKDD International Conference on KnowledgeDiscovery and Data Mining (KDD rsquo10) pp 243ndash252 July 2010

[24] L Munasinghe and R Ichise ldquoLink prediction in social net-works using information flow via active linksrdquo IEICE Transac-tion on Information and Systems vol E96-D no 7 pp 1495ndash1502 2013

[25] Axis httpaxisapacheorgaxis2[26] ns3 httpswwwnsnamorgns3[27] Dataset httpwwwzibinzhengcomtpds

Wireless Communications and Mobile Computing 13

Response-time

JV-PCCTECSSSSRM

CloudRank2

02

025

03

035

04

045

05

055

06

065

07

MA

E

10 25 3015 205k

(a)

Throughput

02

025

03

035

04

045

05

055

06

065

07

MA

E10 15 20 25 305

k

JV-PCCTECSSSSRM

CloudRank2

(b)

Response-time

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 25 3015 205k

JV-PCCTECSSSSRM

CloudRank2

(c)

Throughput

03

035

04

045

05

055

06

065

07

075

08

RMSE

10 15 20 25 305k

JV-PCCTECSSSSRM

CloudRank2

(d)

Figure 7 Impact of neighbor size 119896 comparison with other popular approaches

14 Wireless Communications and Mobile Computing

to select themost trustworthy cloud service of certain type forthe active users based on multidimensional trust evidenceWe also will conduct more experimental investigations thatdeal with the impact of different context changes on serviceselection Moreover we will improve the ranking accuracyof our approaches by exploiting additional techniques (com-bining rating-oriented approaches and ranking-orientedapproaches matrix factorization intelligent optimizationalgorithms etc)

Conflicts of Interest

The author declares that they have no conflicts of interest

Acknowledgments

The work in this paper has been supported by NationalNatural Science Foundation of China (Program no 71501156)and China Postdoctoral Science Foundation (Program no2014M560796) and Shanxi Provincial EducationDepartment(Program no 15JK1679)

References

[1] White PaperMobile CloudComputing Solution Brief AEPONA2010

[2] httpsenwikipediaorgwikiCloudlet[3] Z Zheng X Wu Y Zhang M R Lyu and J Wang ldquoQoS

ranking prediction for cloud servicesrdquo IEEE Transactions onParallel and Distributed Systems vol 24 no 6 pp 1213ndash12222013

[4] Y Pan SDingW Fan J Li and S Yang ldquoTrust-enhanced cloudservice selection model based on QoS analysisrdquo PLoS ONE vol10 no 11 Article ID e0143448 2015

[5] S Ding C-Y Xia K-L Zhou S-L Yang and J S Shang ldquoDeci-sion support for personalized cloud service selection throughmulti-attribute trustworthiness evaluationrdquo PLoS ONE vol 9no 6 Article ID e107821 2014

[6] S K Garg S Versteeg and R Buyya ldquoA framework for rankingof cloud computing servicesrdquo Future Generation ComputerSystems vol 29 no 4 pp 1012ndash1023 2013

[7] H Ma Z Hu K Li and H Zhang ldquoToward trustworthy cloudservice selection a time-aware approach using interval neutro-sophic setrdquo Journal of Parallel and Distributed Computing vol96 pp 75ndash94 2016

[8] A Ahmed and A S Sabyasachi ldquoMobile cloud service selectionusing back propagation neural networkrdquo International Journalof Computer Applications vol 129 no 14 pp 1ndash5 2015

[9] L Guo J Ma Z-M Chen and H-R Jiang ldquoIncorporatingitem relations for social recommendationrdquo Chinese Journal ofComputers vol 37 no 1 pp 219ndash228 2014

[10] Z Yang B Wu K Zheng X Wang and L Lei ldquoA survey ofcollaborative filtering-based recommender systems for mobileinternet applicationsrdquo IEEE Access vol 4 pp 3273ndash3287 2016

[11] Markets and Markets httpwwwmarketsandmarketscomPressReleasesmobile-cloudasp

[12] M Whaiduzzaman A Gani N B Anuar M Shiraz MN Haque and I T Haque ldquoCloud service selection usingmulticriteria decision analysisrdquoTheScientificWorld Journal vol2014 Article ID 459375 10 pages 2014

[13] K Tserpes F Aisopos D Kyriazis and T Varvarigou ldquoArecommender mechanism for service selection in service-oriented environmentsrdquo Future Generation Computer Systemsvol 28 no 8 pp 1285ndash1294 2012

[14] Z U Rehman O K Hussain and F K Hussain ldquoParallelcloud service selection and ranking based on QoS historyrdquoInternational Journal of Parallel Programming vol 42 no 5 pp820ndash852 2014

[15] W-J Fan S-L YangH Perros and J Pei ldquoAmulti-dimensionaltrust-aware cloud service selection mechanism based on evi-dential reasoning approachrdquo International Journal of Automa-tion and Computing vol 12 no 2 pp 208ndash219 2015

[16] X G Wang J Cao and Y Xiang ldquoDynamic cloud serviceselection using an adaptive learning mechanism in multi-cloudcomputingrdquo The Journal of Systems and Software vol 100 pp195ndash210 2015

[17] C T Do N H Tran E-N Huh C S Hong D Niyato andZ Han ldquoDynamics of service selection and provider pricinggame in heterogeneous cloud marketrdquo Journal of Network andComputer Applications vol 69 pp 152ndash165 2015

[18] R Pal and P Hui ldquoEconomic models for cloud service marketspricing and capacity planningrdquo Theoretical Computer Sciencevol 496 pp 113ndash124 2013

[19] A Ezenwoke O Daramola and M Adigun ldquoTowards avisualization framework for service selection in cloud e-marketplacesrdquo in Proceedings of the 2017 IEEE InternationalConference on Web Services (ICWS rsquo17) pp 828ndash835 HonoluluHI USA June 2017

[20] M Tang X Dai J Liu and J Chen ldquoTowards a trust evaluationmiddleware for cloud service selectionrdquo Future GenerationComputer Systems vol 74 pp 302ndash312 2017

[21] A K Dey ldquoUnderstanding and using contextrdquo Personal andUbiquitous Computing vol 5 no 1 pp 4ndash7 2001

[22] X Han L Wang S Park A Cuevas and N Crespi ldquoAlikepeople alike interests A large-scale study on interest similarityin social networksrdquo in Proceedings of the 2014 IEEEACM Inter-national Conference on Advances in Social Networks Analysisand Mining (ASONAM rsquo14) pp 491ndash496 August 2014

[23] R N Lichtenwalter J T Lussier and N V Chawla ldquoNewperspectives and methods in link predictionrdquo in Proceedings ofthe 16th ACM SIGKDD International Conference on KnowledgeDiscovery and Data Mining (KDD rsquo10) pp 243ndash252 July 2010

[24] L Munasinghe and R Ichise ldquoLink prediction in social net-works using information flow via active linksrdquo IEICE Transac-tion on Information and Systems vol E96-D no 7 pp 1495ndash1502 2013

[25] Axis httpaxisapacheorgaxis2[26] ns3 httpswwwnsnamorgns3[27] Dataset httpwwwzibinzhengcomtpds

14 Wireless Communications and Mobile Computing

to select themost trustworthy cloud service of certain type forthe active users based on multidimensional trust evidenceWe also will conduct more experimental investigations thatdeal with the impact of different context changes on serviceselection Moreover we will improve the ranking accuracyof our approaches by exploiting additional techniques (com-bining rating-oriented approaches and ranking-orientedapproaches matrix factorization intelligent optimizationalgorithms etc)

Conflicts of Interest

The author declares that they have no conflicts of interest

Acknowledgments

The work in this paper has been supported by NationalNatural Science Foundation of China (Program no 71501156)and China Postdoctoral Science Foundation (Program no2014M560796) and Shanxi Provincial EducationDepartment(Program no 15JK1679)

References

[1] White PaperMobile CloudComputing Solution Brief AEPONA2010

[2] httpsenwikipediaorgwikiCloudlet[3] Z Zheng X Wu Y Zhang M R Lyu and J Wang ldquoQoS

ranking prediction for cloud servicesrdquo IEEE Transactions onParallel and Distributed Systems vol 24 no 6 pp 1213ndash12222013

[4] Y Pan SDingW Fan J Li and S Yang ldquoTrust-enhanced cloudservice selection model based on QoS analysisrdquo PLoS ONE vol10 no 11 Article ID e0143448 2015

[5] S Ding C-Y Xia K-L Zhou S-L Yang and J S Shang ldquoDeci-sion support for personalized cloud service selection throughmulti-attribute trustworthiness evaluationrdquo PLoS ONE vol 9no 6 Article ID e107821 2014

[6] S K Garg S Versteeg and R Buyya ldquoA framework for rankingof cloud computing servicesrdquo Future Generation ComputerSystems vol 29 no 4 pp 1012ndash1023 2013

[7] H Ma Z Hu K Li and H Zhang ldquoToward trustworthy cloudservice selection a time-aware approach using interval neutro-sophic setrdquo Journal of Parallel and Distributed Computing vol96 pp 75ndash94 2016

[8] A Ahmed and A S Sabyasachi ldquoMobile cloud service selectionusing back propagation neural networkrdquo International Journalof Computer Applications vol 129 no 14 pp 1ndash5 2015

[9] L Guo J Ma Z-M Chen and H-R Jiang ldquoIncorporatingitem relations for social recommendationrdquo Chinese Journal ofComputers vol 37 no 1 pp 219ndash228 2014

[10] Z Yang B Wu K Zheng X Wang and L Lei ldquoA survey ofcollaborative filtering-based recommender systems for mobileinternet applicationsrdquo IEEE Access vol 4 pp 3273ndash3287 2016

[11] Markets and Markets httpwwwmarketsandmarketscomPressReleasesmobile-cloudasp

[12] M Whaiduzzaman A Gani N B Anuar M Shiraz MN Haque and I T Haque ldquoCloud service selection usingmulticriteria decision analysisrdquoTheScientificWorld Journal vol2014 Article ID 459375 10 pages 2014

[13] K Tserpes F Aisopos D Kyriazis and T Varvarigou ldquoArecommender mechanism for service selection in service-oriented environmentsrdquo Future Generation Computer Systemsvol 28 no 8 pp 1285ndash1294 2012

[14] Z U Rehman O K Hussain and F K Hussain ldquoParallelcloud service selection and ranking based on QoS historyrdquoInternational Journal of Parallel Programming vol 42 no 5 pp820ndash852 2014

[15] W-J Fan S-L YangH Perros and J Pei ldquoAmulti-dimensionaltrust-aware cloud service selection mechanism based on evi-dential reasoning approachrdquo International Journal of Automa-tion and Computing vol 12 no 2 pp 208ndash219 2015

[16] X G Wang J Cao and Y Xiang ldquoDynamic cloud serviceselection using an adaptive learning mechanism in multi-cloudcomputingrdquo The Journal of Systems and Software vol 100 pp195ndash210 2015

[17] C T Do N H Tran E-N Huh C S Hong D Niyato andZ Han ldquoDynamics of service selection and provider pricinggame in heterogeneous cloud marketrdquo Journal of Network andComputer Applications vol 69 pp 152ndash165 2015

[18] R Pal and P Hui ldquoEconomic models for cloud service marketspricing and capacity planningrdquo Theoretical Computer Sciencevol 496 pp 113ndash124 2013

[19] A Ezenwoke O Daramola and M Adigun ldquoTowards avisualization framework for service selection in cloud e-marketplacesrdquo in Proceedings of the 2017 IEEE InternationalConference on Web Services (ICWS rsquo17) pp 828ndash835 HonoluluHI USA June 2017

[20] M Tang X Dai J Liu and J Chen ldquoTowards a trust evaluationmiddleware for cloud service selectionrdquo Future GenerationComputer Systems vol 74 pp 302ndash312 2017

[21] A K Dey ldquoUnderstanding and using contextrdquo Personal andUbiquitous Computing vol 5 no 1 pp 4ndash7 2001

[22] X Han L Wang S Park A Cuevas and N Crespi ldquoAlikepeople alike interests A large-scale study on interest similarityin social networksrdquo in Proceedings of the 2014 IEEEACM Inter-national Conference on Advances in Social Networks Analysisand Mining (ASONAM rsquo14) pp 491ndash496 August 2014

[23] R N Lichtenwalter J T Lussier and N V Chawla ldquoNewperspectives and methods in link predictionrdquo in Proceedings ofthe 16th ACM SIGKDD International Conference on KnowledgeDiscovery and Data Mining (KDD rsquo10) pp 243ndash252 July 2010

[24] L Munasinghe and R Ichise ldquoLink prediction in social net-works using information flow via active linksrdquo IEICE Transac-tion on Information and Systems vol E96-D no 7 pp 1495ndash1502 2013

[25] Axis httpaxisapacheorgaxis2[26] ns3 httpswwwnsnamorgns3[27] Dataset httpwwwzibinzhengcomtpds

International Journal of

AerospaceEngineeringHindawiwwwhindawicom Volume 2018

RoboticsJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Active and Passive Electronic Components

VLSI Design

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Shock and Vibration

Hindawiwwwhindawicom Volume 2018

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawiwwwhindawicom

Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Control Scienceand Engineering

Journal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Journal ofEngineeringVolume 2018

SensorsJournal of

Hindawiwwwhindawicom Volume 2018

International Journal of

RotatingMachinery

Hindawiwwwhindawicom Volume 2018

Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Navigation and Observation

International Journal of

Hindawi

wwwhindawicom Volume 2018

Advances in

Multimedia

Submit your manuscripts atwwwhindawicom