Int J Data Sci Anal (2017) 3:35–48DOI 10.1007/s41060-016-0031-0

REGULAR PAPER

Recommendation in heterogeneous information network via dualsimilarity regularization

Jing Zheng1 · Jian Liu1 · Chuan Shi1 · Fuzhen Zhuang2 · Jingzhi Li3 · Bin Wu1

Received: 27 August 2016 / Accepted: 23 October 2016 / Published online: 7 November 2016© Springer International Publishing Switzerland 2016

Abstract Recommender system has caught much attentionfrommultiple disciplines, andmany techniques are proposedto build it. Recently, social recommendation becomes a hotresearch direction. The social recommendationmethods tendto leverage social relations among users obtained from socialnetwork to alleviate data sparsity and cold-start problems inrecommender systems. It employs simple similarity informa-tion of users as social regularization on users. Unfortunately,the widely used social regularization suffers from severalaspects: (1) The similarity information of users only stemsfrom social relations of related users; (2) it only has constrainton users without considering the impact of items for recom-mendation; (3) it may not work well for dissimilar users.To overcome the shortcomings of social regularization, wedesign a novel dual similarity regularization to impose theconstraint on users and items with high and low similaritiessimultaneously. With the dual similarity regularization, wefurther propose anoptimization function to integrate the simi-larity information of users and items under different semanticmeta-paths, and a gradient descend solution is derived tooptimize the objective function. Experiments with differ-

This paper is an extension version of the PAKDD2016 LongPresentation paper “Dual Similarity Regularization forRecommendation” [40].

B Chuan Shishichuan@bupt.edu.cn

1 Beijing Key Lab of Intelligent Telecommunications Softwareand Multimedia, Beijing University of Posts andTelecommunications, Beijing, China

2 Key Laboratory of Intelligent Information Processing,Institute of Computing Technology, Chinese Academyof Sciences, Beijing, China

3 Southern University of Science and Technology, Shenzhen,China

ent meta-paths validate the superiority of integrating muchavailable information, and the experiments conducted onthree real data sets validate the effectiveness of the proposedsolution.

Keywords Social recommendation · Regularization ·Heterogeneous information network

1 Introduction

Recommender system, which aims to recommend items touser personality, has attracted much attention from multipledisciplines. Many entertainment Web sites, such as Netflixand YouTube, have deployed their own recommender sys-tems to increase users’ stickiness. Also many e-commercecompanies, such asAmazon andAlibaba, try to improve theirown recommender systems to increase users’ click conver-sion rate.Recommender systemhas becomepopular in recentyears, and many techniques have been proposed to buildrecommender systems [1,22]. Among them, the low-rankmatrix factorization, as a popular technique, which factorizesuser–item rating matrix into two low-rank user-specific anditem-specific matrices and then utilizes the factorized matri-ces to make further predictions, has shown its effectivenessand efficiency [29].

With the boom of social media, social recommendationhas become a hot research topic. The basic idea of socialrecommendation is to model social network informationas regularization terms to constrain the matrix factoriza-tion framework. Social recommendation generates similarityof users through leveraging rich social relations amongusers, such as friendships in Facebook, following relations inTwitter. Some researchers utilized trust information amongusers [18,19], and some began to use friend relationship

123

36 Int J Data Sci Anal (2017) 3:35–48

among users [20,34] or other types of information [2,5]to enhance recommendation quality. Most of these socialrecommendation methods employ the similarity of users associal regularization to confine similar users under the low-rank matrix factorization framework. Specifically, we canobtain the similarity of users from their social relations as aconstraint term to confine the latent factors of similar usersto be closer. It is reasonable since similar users should havesimilar latent features.

However, we can find that the social regularization usedin social recommendation has following limitations, whichmay lead to poor performance.

(1) The similarity information of users is only generatedfrom social relations of users. On the one hand, in realworld, much more useful information can be utilizedto calculate the similarity of users. For example, we canobtain the similarity of two users through analyzing theirattribute information or rating information. On the otherhand, some users’ social relations are weak, which ishard for similarity analysis. In addition, the character-istics of items connecting to two users who have closesocial relation may be very different, so that we cannotsimply employ the similarity generated from their socialrelation to constrain the latent factor of these two users.Therefore, it is more reasonable to integrate more usefulinformation of users for users’ similarity analysis.

(2) The social regularization only has constraint on users. Infact, we can obtain the similarity information of items aswell and then we can use it to impose constraint on thelatent factors of items. When the information of usersis sparse, it is difficult to obtain users’ similarity infor-mation. In this condition, we cannot employ the socialrecommendation. On the other hand, available informa-tion of items can be collected to analyze the similarityof items, which can be used to confine the latent factorsof items for better recommendation.

(3) The social regularization may be less effective for dis-similar users. When two users are similar, the socialregularization takes effect and forces the latent factorsof these two users to be close. However, when two usersare not similar, the social regularizationmay fail and stillforces the latent factors of these two ones to be close,which is inconsistent with their social relation in real-ity. The analysis and experiments in Sect. 3 validate thispoint.

In order to address the above limitations we discussed,we propose a dual similarity regularization-based recom-mendation method (called DSR) in this paper. Inspired bythe success of heterogeneous information network (HIN) inmany applications, we organize objects and relations in arecommender system as a HIN, which can integrate vari-

ous information, including interactions between users anditems, social relations among users and attribute informa-tion of users and items. Based on the HIN, we generaterich similarity information on users and items by settingproper meta-paths. Furthermore, we propose a new dual sim-ilarity regularization which can constrain the latent featuresof users and items with arbitrary similarity simultaneously.With the similarity regularization, DSR adopts a new opti-mization objective to integrate the similarity information ofusers and items. Then, we derive its solution to learn theweights of different similarities. The experiments on threereal data sets (i.e., Douban-Movie, Douban-Book and Yelp)show that DSR always performs better compared to socialrecommendation and twoHIN-based recommendationmeth-ods (i.e.,HeteCFandHeteMF).Moreover,DSRalso achievesthe best performance for cold-start users and items due to thedual similarity regularization on users and items. The maincontributions of the paper are summarized as follows:

(1) We design a novel dual similarity regularization, whichimposes constraints on users and items with high and lowsimilarities simultaneously. The dual similarity regular-ization would force the latent factors of two dissimilarobjects to be far and force the latent factors of two simi-lar ones to be close, which can avoid the weakness of thesocial regularization for two dissimilar users.

(2) We propose a unified HIN-based social recommendationmethod with dual similarity regularization, called DSRfor short, which is based on the basic matrix factorizationframework. DSR can integratemultiple types of informa-tion in HIN to enhance recommendation performance. Inaddition, DSR can obtain more accurate integrated sim-ilarity information of objects through flexible weightsassigned to similarities of users or items under differentsemantic paths.

(3) Extensive experimentation conducted on three real datasets demonstrates the effectiveness of DSR. In particu-lar, in terms of providing recommendation to cold-startobjects (i.e., users, items or both of them), DSR canachieve comparable performance, which shows the supe-riority of our proposed method in alleviating cold-startproblem.

The remainder of this paper is organized as follows. Wereview the related work in Sect. 2. Next, we analyze the limi-tations of social recommendation in Sect. 3 and introducethe rich similarity information of users and items gener-ated from HIN in Sect. 4. Then, we propose the similarityregularization and the DSR model in Sect. 5. Experimentalresults are reported in Sect. 6, and the conclusion is drawn inSect. 7.

123

Int J Data Sci Anal (2017) 3:35–48 37

2 Related work

Recommender system aims to tackle the information over-load problem, and many recommender methods have beenpresented to improve recommender performance. Traditionalrecommender methods usually utilize the user–item ratinginformation for recommendation [7,21]. To further enhancethe recommendation performance, external information suchas the attributes of users and items is considered in manymethods [8,11]. Collaborative filtering [10] as one of themost successful approaches, which utilizes information ofusers preferences to predict items a target user might like,can be grouped into two types: memory-based methods andmodel-based methods. However, most of them cannot dealwith the cold-start problem which is a situation that recom-mender system fails to draw recommendation for new items,users or both.

With the prevalence of social media, social recommen-dation has attracted many researchers. Ma et al. [19] fuseduser–item matrix with users’ social trust networks. Bernardiet al. [3] proposed a recommendation algorithm to solve thecold-start problem with social tags. In [20], the social regu-larization ensures the latent feature vectors of two friendswith similar tastes to be closer. Yang et al. [34] inferredcategory-specific social trust circles from available ratingdata combined with friend relations. Moreover, Zhang etal. [38] proposed a novel BiFu scheme for the cold-startproblem based on the bi-clustering and fusion techniquesin large-scale social recommender systems. Consideringthat users may be affiliated with multiple groups in onlinesocial networks, Yang et al. [33] utilized social groups withricher information to improve the recommendation perfor-mance.

To further improve recommendation performance, moreand more researchers have been aware of the importanceof heterogeneous information network (HIN), in whichobjects are of different types and links among objectsrepresent different relations [25]. Zhang et al. [39] investi-gated the problem of recommendation over heterogeneousnetwork and proposed a random walk model to estimatethe importance of each object in the heterogeneous net-work. Considering heterogeneous network constructed bydifferent interactions of users, Jamali and Lakshmanan [13]proposed HETEROMF to integrate a general latent factorand context-dependent latent factors. Yu et al. [36,37] pro-posed an implicit feedback recommendation model withextracted latent features from HIN. Moreover, they [35] pro-posed HeteMF through combining rating information anditems’ similarities derived from meta-paths in HIN. Luo etal. [17] proposed a collaborative filtering-based social rec-ommendation method, called HeteCF, using heterogeneousrelations. Recently, Shi et al. [27] proposed the SemRecon HIN to improve recommendation performance through

flexibly integrating information with the help of weightedmeta-paths. Furthermore, the SimMF proposed by Shi etal. [26] integrates heterogeneous information via flexibleregularization of users and items for better recommenda-tion.

With the development of recommender systems, muchmore additional information and learning tools have beenexploited to improve the recommendation quality. Guo etal. [9] proposed an effective algorithm called LBMF toexplore review texts and rating scores simultaneously, whichdirectly correlate user and item latent dimensions witheach word in review texts and ratings. Wu et al. [32] pre-sented a novel method for top-N recommendation, calledcollaborative denoising auto-encoder (CDAE), which uti-lized the idea of denoising auto-encoders. Cao et al. [31]noticed not only the heterogeneity but also the couplingsbetween objects. They proposed the non-IID recommen-dation theories and systems [6]. Considering the non-IID context, Li et al. [16] presented a coupled matrixfactorization framework, which integrated user couplingsand item couplings into the basic matrix factorizationmodel.

Contemporary recommender methods tend to considersingle type or two types of information, even for heteroge-neous methods. For example, HeteMF proposed by Yu et al.only takes similarity of items into account to improve recom-mendation performance. Following the existing HIN-basedmethods, we proposed a novel recommendation methodcalled DSR in HIN, where we can obtain rich useful infor-mation of users and items. In order to make full use of theavailable information, DSR flexible integrates the informa-tion of users and items under different semantic paths toimprove the recommendation performance. Ma et al. pro-posed SoMF for social recommendation and designed thesocial regularization to ensure the latent feature vectors oftwo friends with similar tastes to be closer; however, it maynot work well for dissimilar users. Compared with the socialregularization, the dual similarity regularizationwe proposedimposes the constraint not only on similar users but also ondissimilar users.

3 Limitations of social recommendation

Recently, with the increasing popularity of social media,there is a surge of social recommendations which lever-age rich social relations among users to improve recom-mendation performance. Ma et al. [20] first proposed thesocial regularization to extend low-rank matrix factoriza-tion, and then, it is widely used in a lot of work [17,35].A basic social recommendation method is illustrated asfollows:

123

38 Int J Data Sci Anal (2017) 3:35–48

minU,V

J =1

2

m∑

i=1

n∑

j=1

Ii j(Ri j −UiV

Tj

)2

+ α

2

m∑

i=1

m∑

j=1

SU (i, j)‖Ui −Uj‖2

+ λ1

2(‖U‖2 + ‖V ‖2),

(1)

where m × n rating matrix R depicts users’ ratings on nitems, Ri j is the score user i gives to item j . Ii j is an indi-cator function which equals to 1 if user i rated item j andequals to 0 otherwise. U ∈ R

m×d and V ∈ Rn×d , where

d << min(m, n) is the dimension number of latent fac-tor. Ui is the latent vector of user i derived from the i throw of matrix U while Vj is the latent vector of item jderived from the j th row of V . SU is the similarity matrixof users and SU (i, j) denotes the similarity of user i anduser j . ‖ · ‖2 is the Frobenius norm. Particularly, the secondterm is the social regularization which is defined as follows:

SocReg =1

2

m∑

i=1

m∑

j=1

SU (i, j)‖Ui −Uj‖2. (2)

As a constraint term in Eq. (1), SocReg forces the latentfactors of two users to be close when they are very similar.However, it may have two drawbacks.

• The similarity information may be simple. In socialrecommendation, the similarity information of users isusually generated from rating information or social rela-

tions and only one type of similarity information isemployed. However, we can obtain much rich similar-ity information of users and items from various ways,such as rich attribute information and interactions.

• The constraint term may not work well when two usersare not very similar. The minimization of optimizationobjective should force the latent factors of two similarusers to be close. But we note that when two users arenot similar (i.e., SU (i, j) is small), SocRegmay still forcethe latent factors of these two users to be close. In fact,these two users are dissimilar which means their latentfactors should have a large distance.

In order to uncover the limitations of social regularization,we apply the model detailed in Eq. (1) to conduct four exper-iments each with different levels of similarity information(None, Low, High, All). None denotes that we do not uti-lize any similarity information in the model (i.e., α=0 in themodel), Low denotes that we utilize the bottom 20% users’similarity information generated in the model, High is thatof top 20%, and All denotes we utilize all users’ similarityinformation. The Douban-Movie data set detailed in Table 1is employed in the experiments, and we report MAE andRMSE (defined in Sect. 6.2) in Fig. 1.

The results of Low, High and All are better than thatof None, which implies social regularization really worksin the model. However, in terms of performance improve-ment compared to None, Low does not improve as much asHigh and All do. The above analysis reveals that the social

Table 1 Statistics of Doubanand Yelp data set

Data sets Relations (A–B) Number of A/B/A–B Ave. degrees of A/B

Douban-Movie User–Movie 3022/6971/195,493 64.69/28.04

User–User 779/779/1366 1.75/1.75

User–Group 2212/2269/7054 3.11/3.11

User–Location 2491/244/2491 1.00/10.21

Movie–Director 3014/789/3314 1.09/4.20

Movie–Actor 5438/3004/15,585 2.87/5.19

Movie–Type 6787/36/15,598 2.29/433.28

Douban-Book User–Book 13,024/22,347/792,062 60.82/35.44

User–User 12,748/12,748/169,150 13.27/13.27

User–Group 13,024/2936/1,189,271 91.31/405.07

User–Location 10592/453/10592 1.00/23.38

Book–Publisher 21,773/1815/21,773 1.00/12.00

Book–Author 21,907/10,805/21,907 1.00/2.03

Book–Year 21,192/64/21,192 1.00/331.13

Yelp User–Business 14,085/14,037/194,255 4.6/20.7

User–User 9581/9581/150,532 10.0/10.0

Business–Category 14,037/575/39,406 2.8/73.9

Business–Location 14,037/62/14,037 1.0/236.1

123

Int J Data Sci Anal (2017) 3:35–48 39

None Low High All0.6

0.61

0.62

0.63

0.64

0.65M

AE

different similarity of usersNone Low High All

0.73

0.75

0.77

0.79

0.81

0.83

RM

SE

Fig. 1 Limitations of social regularization

regularization may not work well in recommender modelswhen users are with low similarity.

4 Rich similarity generated from HIN

An information network represents an abstraction of the realworld, focusing on the objects and the interactions amongthem. The information network analysis has become a hotspot in data mining and information retrieval fields in thepast decades, whose base idea is to mine link relations frommassive data in network to find the hidden patterns of differ-ent objects.

A number of contemporary recommendation methodsmodel their problems in homogeneous information networks,which ignore the heterogeneity of objects andmerely containone type of relations among objects with the same type, suchas the friendship network [15]. Traditional social recommen-dations only consider the constraint of users with their socialrelations in an information network. However, most of therecommender systems contain multiple types of objects andrelation links. It is hard to model them as homogeneous net-works while rich similarity information on users and items

can be generated in a heterogeneous information network.A heterogeneous information network [12] is a specialinformation network with multiple types of entities and rela-tions. Compared to traditional homogeneous informationnetworks, the HIN can integrate much more available infor-mation effectively and contain richer semantics in objectsand links. Moreover, the homogeneous information networkis usually constructed from single data domain, while theHIN can fuse information across domains. For example,users always tend to participate in different networks, suchas YouTube, Facebook and Twitter.

Recently, the HIN has aroused general interests and manyrecommendation tasks have been exploited in such networksfor better recommendation. Figure 2a shows a typical HINextracted from a movie recommender system. The HIN con-tains multiple types of objects, e.g., users (U), movies (M),groups (G) and actors (A). Also, it illustrates different kindsof relations among objects, e.g., rating information, attributeinformation and social relation.

Two types of objects in aHIN can be connected via variousmeta-path [30], which is a composite relation connectingthese two types of objects. Ameta-pathP is a path defined on

a schema S = (A,R), and is denoted in the form of A1R1−→

A2R2−→ · · · Rl−→ Al+1 (abbreviated as A1A2 . . . Al+1), which

defines a composite relation R = R1 ◦ R2 ◦ · · · ◦ Rl betweentype A1 and Al+1, where ◦ denotes the composition operatoron relations. As an example in Fig. 2a, users can be con-nected via “User–User” (UU), “User–Movie–User” (UMU)and so on. Different meta-paths denote different semanticrelations, e.g., the UU path means that users have social rela-tions, while the UMU path means that users have watchedthe same movies. Therefore, we can evaluate the similar-ity of users (or movies) based on different meta-paths. Forexample, for users, we can consider UU, UGU, UMU, etc.Similarly, meaningfulmeta-paths connectingmovies includeMAM and MDM.

Several path-based similarity measures have been pro-posed to evaluate the similarity of objects under givenmeta-path in HIN [14,24,30]. Sun et al. [30] implement

User Movie1 - 5

Actor

Type

(a) (b) (c)

Director

Group

Location

User Book1 - 5

Author

Year

Publisher

Group

Location

User Business1 - 5

Location

Category

Fig. 2 Network schema of HIN examples. a Douban-Movie, b Douban-Book, c Yelp

123

40 Int J Data Sci Anal (2017) 3:35–48

the PathSim to measure the similarity of objects with sametype based on symmetric paths. Lao et al. [14] propose apath-constrained random walk (PCRW) model to measurethe similarity of two objects in a labeled directed graph. Shiet al. [28] presented a semantic-based recommendation sys-tem HeteRecom, which utilizes the semantics informationof meta-path to evaluate the similarities between movies.The HeteSim [24] is proposed to measure the relatednessof heterogeneous objects based on an arbitrary meta-path.Burke et al. [4] proposed an approach for recommendationthrough incorporatingmultiple relations in aweighted hybridsystem.

We assume that S(p)U denotes similarity matrix of users

under meta-path P(p)U connecting users, and S(p)

U (i, j)

denotes the similarity of users i and j under the path P(p)U .

Similarly, S(q)I denotes similarity matrix of items under the

path P(q)I connecting items, and S(q)

I (i, j) denotes the simi-larity of items i and j . Since users (or items) have differentsimilarities under variousmeta-paths, we combine their simi-larities on all paths through assigning weights on these paths.For users and items, we define SU and SI to represent thesimilaritymatrix of users and items on allmeta-paths, respec-tively.

SU =|PU |∑

p=1

w(p)U S(p)

U ,

SI =|PI |∑

q=1

w(q)I S(q)

I ,

(3)

where w(p)U denotes the weight of meta-path P(p)

U among all

meta-pathsPU connecting users, andw(q)I denotes theweight

of meta-path P(q)I among all meta-paths PI connecting

items.

5 Matrix factorization with similarityregularization

In this section, we propose our dual similarity regularization-basedmatrix factorizationmethodDSR and infer its learningalgorithm. We first review the basic low-rank matrix factor-ization method. Then, we introduce dual similarity regular-ization which imposes the constraint on users and items withhigh and low similarity. Based on the basic matrix factor-ization framework and the regularization we designed, wepropose the dual similarity regularization for recommenda-tion (called DSR). Finally, we infer the learning algorithmof DSR and show the unified optimization algorithm frame-work.

5.1 Low-rank matrix factorization

The low-rank matrix factorization has been widely studiedin recommendation system. Its basic idea is to factorize theuser–item rating matrix R into two matrices (U and V ) rep-resenting users’ and items’ distributions on latent semantic,respectively. Then, the rating prediction can bemade throughthese two specificmatrices. Assuming anm×n ratingmatrixR to be m users’ ratings on n items, this approach mainlyminimizes the objective function J as follows.

minU,V

J =1

2

m∑

i=1

n∑

j=1

Ii j(Ri j −UiV

Tj

)2

+ λ1

2(‖U‖2 + ‖V ‖2),

(4)

where Ii j is the indicator function that is equal to 1 if useri rated item j and equal to 0 otherwise. U ∈ R

m×d andV ∈ R

n×d , where d is the dimension number of latent factorsand d << min(m, n). Ui is a row vector derived from i throw of matrix U , and Vj is a row vector derived from j throw of matrix V . λ1 represents the regularization parameter.

In summary, the optimization problem minimizes thesum-of-squared-errors objective function with quadratic reg-ularization terms which aim to avoid over-fitting problem.This problem can be effectively solved by a simple stochas-tic gradient descent technique.

5.2 Similarity regularization

Due to the limitations of social regularization, we design anew similarity regularization to constrain users and itemssimultaneously with much similarity information of usersand items. The basic idea of similarity regularization is thatthe distance of latent factors of two users (or items) should benegatively correlated with their similarity, which means twosimilar users (or items) should have a short distance whiletwo dissimilar ones should have a long distance with theirlatent factors.Wenote that theGaussian functionmeets aboverequirement and the range of it is [0,1], which is the samewith the range of similarity function. Following the idea, wedesign a similarity regularization on users as follows:

SimRegU =1

8

m∑

i=1

m∑

j=1

(SU (i, j) − e−γ ‖Ui−Uj‖2

)2, (5)

whereγ controls the radial intensity ofGaussian function andthe coefficient 1

8 is convenient for deriving the learning algo-rithm. This similarity regularization can enforce constrainton both similar and dissimilar users. In addition, the simi-larity matrix SU can be generated from social relations orthe above HIN. Similarly, we can also design the similarityregularization on items as follows:

123

Int J Data Sci Anal (2017) 3:35–48 41

SimRegI =1

8

n∑

i=1

n∑

j=1

(SI (i, j) − e−γ ‖Vi−Vj‖2

)2. (6)

5.2.1 The proposed DSR model

We propose the dual similarity regularization for recommen-dation (called DSR) through adding the similarity regular-ization on users and items into low-rank matrix factorizationframework. Specifically, the optimization model is proposedas follows:

minU,V,wU ,w I

J = 1

2

m∑

i=1

n∑

j=1

Ii j(Ri j −UiV

Tj

)2

+λ1

2(‖U‖2 + ‖V ‖2) + λ2

2

(‖wU‖2

+‖w I‖2)

+ αSimRegU + βSimRegI

s.t.|PU |∑

p=1

w(p)U = 1,w(p)

U ≥ 0

|PI |∑

q=1

w(q)I = 1,w(q)

I ≥ 0, (7)

where α and β control the ratio of similarity regularizationterm on users and items, respectively.

5.3 The learning algorithm

The learning algorithm of DSR can be divided into two steps.

(1) Optimize the latent factor matrices of users and items(i.e., U , V ) with the fixed weight vectors wU =[w

(1)U ,w

(2)U , . . . ,w

(|PU |)U

]Tand w I =

[w

(1)I ,w

(2)I , . . . ,

w(|PI |)I

]T.

(2) Optimize the weight vectors wU and w I with the fixedlatent factor matricesU and V . Through iteratively opti-mizing these two steps, we can obtain the optimalU , V ,wU and w I .

5.3.1 Optimize U and V

With the fixed wU and w I , we can optimize U and V byperforming stochastic gradient descent.

∂J∂Ui

=n∑

j=1

Ii j(Ui V

Tj − Ri j

)Vj + λ1Ui

+α

m∑

j=1

γ[(

SU (i, j) − e−γ ‖Ui−Uj‖2)

e−γ ‖Ui−Uj‖2 (Ui −Uj )], (8)

∂J∂Vj

=m∑

i=1

Ii j(Ui V

Tj − Ri j

)Ui + λ1Vj

+β

n∑

i=1

γ[(

SI (i, j) − e−γ ‖Vi−Vj‖2)

e−γ ‖Vi−Vj‖2 (Vi − Vj )]. (9)

5.3.2 Optimize wU and w I

With the fixed U and V , the minimization of J with respectto wU and w I is a well-studied quadratic optimization prob-lem with nonnegative bound. We can use the standard trustregion reflective algorithm to updatewU andw I at each iter-ation. We can simplify the optimization function of wU asthe following standard quadratic formula:

minwU

1

2wTU HUwU + f TU wU

s.t.|PU |∑

p=1

w(p)U = 1,w(p)

U ≥ 0.(10)

Let S(i j)UU = S(i)

U � S( j)U (1 ≤ i, j ≤ |PU |). Here HU is a

|PU | × |PU | symmetric matrix as follows:

HU (i, j) =⎧⎨

⎩

α4

(∑∑S(i j)UU

)i = j, 1 ≤ i, j ≤ |PU |

α4

(∑∑S(i j)UU

)+ λ2 i = j, 1 ≤ i, j ≤ |PU |,

� denotes the dot product. fU is a column vector with length|PU |, which is calculated as follows:

fU (p) = −α

4

m∑

i=1

m∑

j=1

S(p)U (i, j)e−γ ‖Ui−Uj‖2 .

Similarly, we can also infer the optimization function of w I .Algorithm 1 shows the optimization algorithm framework.

6 Experiments

In this section, we conduct experiments to validate the effec-tiveness of DSR and further explore the cold-start problem.

6.1 Data preparation

Although many public data sets are available, some of themmay not contain enough objects attributes information orsocial relation information. In order to get more availableinformation of objects and ensure the generality of ourmethod, we use three real data sets in our experiments. Two

123

42 Int J Data Sci Anal (2017) 3:35–48

Algorithm 1 Alogrithm of DSRRequire:

G: heterogeneous information networkPU , PI : meta-path sets of users and itemsη: learning rate for gradient descentα, β, λ1, λ2: see details in above sectionε: convergence tolerance

Ensure:U, V : the latent factor of users and itemswU ,w I : the weights of meta-paths on users and items

1: for P(p)U ∈ PU and P(q)

I ∈ PI do

2: Calculate the similarity of users and items (S(p)U and S(q)

I )3: end for4: Initialize U, V,wU ,w I5: repeat6: Uold := U ,Vold := V7: Calculate ∂J

∂U , ∂J∂V

8: Update U := U − η ∗ ∂J∂U ,V := V − η ∗ ∂J

∂V9: Update wU ,w I with quadratic optimization10: until ‖U −Uold‖2 + ‖V − Vold‖2 < ε

data sets are from Douban,1 a well-known social media Website in China. In this Web site, data in two domains includingmovie and book are collected and extracted for our exper-iments. The data set in movie domain includes 3022 usersand 6971 movies with 195,493 ratings ranging from 1 to 5,while the data set in book domain comprises 792,062 ratings(scales 1–5) by 13,024 users on 22,347 books. And anotherreal data set is employed from Yelp,2 a famous user reviewWeb site inAmerica, which includes 14,085 users and 14,037movies with 194,255 ratings ranging from 1 to 5.

The description of three data sets is shown in Table 1, andtheir network schemas are shown in Fig. 2. The Douban-Movie data set has sparse social relationship with denserating information, while the Douban-Book data set and theYelp data set have dense social relationships with sparserating information, which leads to different experimentalresults.

6.2 Comparison methods and metrics

In order to validate the effectiveness of DSR, we compareit with following representative methods. Besides the classi-cal social recommendation method SoMF, the experimentsalso include three recent HIN-based methods HeteCF, Het-eMF and CMF. In addition, we include the revised version ofSoMF with similarity regularization (i.e., SoMFSR) to vali-date the effectiveness of similarity regularization.

• UserMean. It employs a user’s mean rating to predict themissing ratings directly.

1 http://www.douban.com/.2 http://www.yelp.com/.

• ItemMean. It employs an item’s mean rating to predictthe missing ratings directly.

• ItemCF [22]. Sarwar et al. proposed the item-based col-laborative filtering recommendation algorithm.

• PMF [21]. Salakhutdinov and Minh proposed the basiclow-rank matrix factorization method for recommenda-tion.

• SoMF [20]. Ma et al. proposed the social recommenda-tion method with social regularization on users.

• HeteCF [17]. Luo et al. proposed the social collaborativefiltering algorithm using heterogeneous relations.

• HeteMF [35]. Yu et al. proposed the HIN-based recom-mendation method through combining user ratings anditems’ similarity matrices.

• CMF [16]. Li et al. proposed the coupled matrix fac-torization recommendation method integrating user cou-plings and item couplings into the basic MF model.

• SoMFSR. It adapts SoMF through only replacing thesocial regularization with the similarity regularizationSimRegU∫ .

For Douban-Movie data set, we utilize 7 meta-paths foruser (i.e., UU, UGU, ULU, UMU, UMDMU, UMTMU andUMAMU) and 5 meta-paths for item (i.e., MTM, MDM,MAM, MUM and MUUM). For Douban-Book data set, weutilize 7 meta-paths for user (i.e., UU, UGU, ULU, UBU,UBPBU, UBABU and UBYBU) and 7 meta-paths for item(i.e., BUB,BPB,BAB,BYB,BUUB,BUGUBandBULUB).For Yelp data set, we utilize 4 meta-paths for user (i.e., UU,UBU, UBCBU and UBLBU) and 4 meta-paths for item (i.e.,BUB, BCB, BLB and BUUB). HeteSim [24] is employedto evaluate the object similarity based on above meta-paths.These similarity matrices are fairly utilized for HeteCF, Het-eMF and DSR.

We set γ = 1, α = 10 and β = 10 through parameterexperiments on Douban-Movie and Douban-Book data sets.In the experiments on Yelp data set, we set the parametersγ = 1, α = 10 and β = 10. Meanwhile, optimal parametersare set for other models in the experiments. We use meanabsolute error (MAE) and root-mean-square error (RMSE)to evaluate the performance of rating prediction:

MAE =∑

(u,i)∈R |Ru,i − R̂u,i ||R| , (11)

RMSE =√∑

(u,i)∈R(Ru,i − R̂u,i )2

|R| , (12)

where R denotes the whole rating set, Ru,i denotes the ratinguser u gave to item i , and R̂u,i denotes the rating user ugave to item i as predicted by a certain method. R̂u,i canbe calculated by UiV T

j . A smaller MAE or RMSE means abetter performance.

123

Int J Data Sci Anal (2017) 3:35–48 43

6.3 Effectiveness experiments

For Douban data sets of movie and book, we use differentratios (80, 60, 40%) of data as training sets and the rest of thedata set for testing. For example, the training ratio 80%meansthatwe randomly select 80%of rating information as trainingdata while the other 20% of rating information as testing datain our experiments. Considering the sparse density of Yelpdata set, we use 90, 80, 70% of data as training sets and therest as test sets for Yelp data set. The random selection iscarried out 10 times independently. We report the averageresults on Douban and Yelp data sets in Table 2. From theresults, we have following observations.

• It is clear that four HIN-based methods (DSR, HeteCF,HeteMF and CMF) all achieve significant performanceimprovements compared to PMF, UserMean, ItemMean,ItemCF and SoMF. With more additional informationto confine the latent factors of users or items, the per-formance can be improved to some extent. It impliesthat integrating heterogeneous information is a promisingway to improve recommendation performance. Further-more, we can see that DSR and CMF always performbetter than HeteMF and SoMF, which indicates thatfusing richer information can improve the recommen-dation quality better. We can also find that CMF andDSR always achieve the best performances. It alsoshows the benefits of exploring the feature correlation inCMF.

• DSR outperforms most of the baselines, including threeHIN-based methods (DSR, HeteCF and HeteMF), whichindicates that the dual similarity regularization on usersand items may be more effective than traditional socialregularization. It can be further confirmed by the betterperformance of SoMFSR over SoMF.

• On three real-world data sets, SoMFSR always per-forms better than SoMF. Although the superiority ofSoMFSR over SoMF is not significant, the improvementis achieved on the very weak social relations in Douban-Movie data set. Note that, when training data becomessparser (for example, when the training ratio is 70%in Yelp data set), SoMFSR performs better or not badthan PMF, while SoMF performs worse than PMF. Theseobservations further confirmed the superiority of the dualsimilarity regularization.

• Observing the results on three data sets, we can alsofind that DSR has better performance improvement forless training data. On Douban-Movie data set, four HIN-based methods (DSR, HeteCF, HeteMF and CMF) andthe SoMF all have better results than PMF. However,on Yelp data set, HeteCF and SoMF improve perfor-mances very limitedly compared with PMF. What is

worse, the results of HeteMF are even worse than PMF.But we can see that DSR always achieves significant per-formance improvement compared with PMF. It revealsthat DSR has the potential to alleviate the cold-startproblem.

• CMF takes constraints on users and items through utiliz-ing similarity of users and items.When there are abundanttraining data, CMF can achieve relatively better perfor-mance than DSR and other HIN-based methods, since itcompromises non-IIDcontext into basicmatrix factoriza-tion framework which is different from other HIN-basedmethods which do not consider the coupling relation-ships between two objects. However, when training dataare sparse (e.g., 70% Yelp training data), the experimen-tal improvement in CMF is less than DSR, because DSRcan integrate not only the attribute information of objectsbut also the social information and rating information ofthem.

6.4 Study on cold-start problem

Cold start is an important issue in recommender system. It isgiving recommendations to novel users who have no or littlepreference on any items, or recommending items that littleor no user of the community has seen yet [23]. To validatethe superiority of DSR on cold-start problem, we run PMF,SoMF, HeteCF, HeteMF, DSR on Douban-Movie data setwith 40% training ratio.We set four levels of users, including:three types of cold-start users with various numbers of ratedmovies (e.g., [0,8] denotes users rated nomore than 8moviesand “All” means all users in Fig. 3). We conduct similarexperiments on cold-start items and users&items (users anditems are both cold-start).

The experiments are shown in Fig. 3. Once again, we findthat three HIN-based methods all are effective for cold-startusers and items.Moreover, DSR always has the highestMAEandRMSE improvement on almost all conditions, due to dualsimilarity regularization on users and items. It is reasonablesince the DSR method takes much constraint information ofusers and items into account which would play a crucial rolewhen there’s little available information of users or items.

6.5 Impact of different meta-paths

In this section, we study the impact of different meta-pathson the Douban-Movie data set. In each experiment, we runour DSR with the similarity information from only onemeta-path. UU∗ and MM∗ represent the integrated similar-ity matrix of users and items on all meta-paths, respectively.Optimal parameters are set for all experiments. The experi-ment results are shown in Fig. 4.

123

44 Int J Data Sci Anal (2017) 3:35–48

Table2

Effectiv

enessexperimentson

DoubanandYelp(the

improvem

entisbasedon

PMF)

Dataset

Training

Metrics

PMF

UserM

ean

Item

Mean

Item

CF

SoMF

HeteC

FHeteM

FCMF

SoMFSR

DSR

Dou

ban-Movie

80%

MAE

0.64

440.69

540.62

840.70

070.63

960.61

010.59

410.56

840.63

360.58

56

Improve

−7.92%

2.47

%−8

.73%

0.73

%5.32

%7.79

%11

.80%

1.68

%9.12

%

RMSE

0.81

510.86

580.79

280.87

810.80

980.76

570.75

200.71

940.80

000.73

79

Improve

−6.23%

2.73

%−7

.73%

0.64

%6.05

%7.73

%11

.74%

1.85

%9.46

%

60%

MAE

0.67

800.69

670.63

700.70

220.66

960.63

170.60

560.57

320.66

480.59

46

Improve

−2.76%

6.05

%−3

.57%

1.25

%6.84

%10

.68%

15.46%

1.96

%12

.31%

RMSE

0.85

690.86

870.81

350.88

090.84

450.79

010.76

650.72

540.83

580.74

83

Improve

−1.37%

5.07

%2.80

%1.45

%7.80

%10

.56%

15.35%

2.46

%12

.68%

40%

MAE

0.73

640.70

090.66

290.70

640.72

450.67

620.62

550.58

330.71

410.60

92

Improve

4.83

%9.99

%4.08

%1.63

%8.18

%15

.07%

20.79%

3.03

%17

.28%

RMSE

0.92

210.87

470.87

470.88

630.90

580.84

040.78

910.73

730.89

500.76

29

Improve

5.14

%5.13

%3.88

%1.76

%8.86

%14

.42%

20.04%

2.94

%17

.27%

Dou

ban-Boo

k80

%MAE

0.57

290.62

070.60

550.62

440.57

130.56

670.57

810.57

290.57

090.55

78

Improve

−8.34%

−5.70%

−8.99%

0.28

%1.07

%−0

.90%

0.01

%0.35

%2.64

%

RMSE

0.73

220.78

990.75

170.79

120.72

590.71

740.73

520.72

350.72

660.70

01

Improve

−7.89%

−2.67%

−8.05%

0.86

%2.02

%−0

.42%

1.19

%0.77

%4.38

%

60%

MAE

0.58

820.62

440.60

920.62

730.58

860.57

750.59

930.58

780.58

720.56

61

Improve

−6.16%

−3.58%

−6.65%

−0.08%

1.81

%−1

.88%

0.07

%0.17

%3.76

%

RMSE

0.75

100.79

910.75

880.79

670.74

770.73

230.76

080.74

310.74

680.71

08

Improve

−6.40%

0.44

%−6

.09%

−0.94%

2.50

%−1

.30%

1.05

%0.57

%5.35

%

40%

MAE

0.61

490.63

300.62

320.63

440.61

710.59

270.64

250.65

840.61

560.57

88

Improve

−2.95%

−1.35%

−3.16%

−0.37%

3.60

%−4

.49%

−7.07%

−0.13%

5.87

%

RMSE

0.78

440.82

090.79

440.81

110.78

580.75

460.81

270.82

920.78

360.72

93

Improve

−4.64%

−1.27%

−3.44%

−0.17%

3.81

%−3

.60%

−5.71%

0.11

%7.02

%

Yelp

90%

MAE

0.84

750.95

430.88

220.94

430.84

600.84

610.89

600.80

990.84

590.81

58

Improve

−12.60

%−4

.09%

−11.42

%0.18

%0.17

%−5

.72%

4.44

%0.18

%3.74

%

RMSE

1.07

961.31

381.21

061.29

321.07

721.07

731.12

721.03

431.07

721.03

69

Improve

−21.69

%−1

2.13

%−1

9.78

%0.22

%0.21

%−4

.41%

4.20

%0.22

%3.95

%

80%

MAE

0.85

280.96

210.89

310.96

580.85

270.85

280.89

070.81

970.85

260.82

06

Improve

−12.82

%−4

.72%

−13.25

%0.01

%0.00

%−4

.44%

3.88

%0.01

%3.78

%

RMSE

1.08

501.32

551.23

041.33

721.08

491.08

501.11

951.04

671.08

481.04

13

Improve

−22.17

%−1

3.40

%−2

3.24

%0.01

%0.00

%−3

.18%

3.53

%0.02

%4.03

%

70%

MAE

0.85

760.97

060.90

620.98

600.85

750.85

760.89

760.82

980.85

750.82

50

Improve

−13.17

%−5

.67%

−26.65

%0.01

%0.00

%−4

.66%

3.24

%0.01

%3.80

%

RMSE

1.08

941.33

951.25

471.37

981.09

361.08

941.13

131.05

921.08

941.04

61

Improve

−22.96

%−1

5.17

%−2

6.65

%−0

.39%

0.00

%−3

.85%

2.77

%0.00

%3.97

%

Boldvalues

representthe

bestresults

ineach

groupof

experiments

123

Int J Data Sci Anal (2017) 3:35–48 45

[0,8] [9,20] [21,) All0

5%

10%

15%

20%

# rating of users

MA

E im

pro

vem

ent SoMF

HeteCFHeteMFDSR

[0,8] [9,20] [21,) All−5%

0

5%

10%

15%

20%

# rating of items

MA

E im

pro

vem

ent

[0,8] [9,20] [21,) All−5%

0

5%

10%

15%

20%

# rating of users&items

MA

E im

pro

vem

ent

[0,5] [6,20] [21,) All

0

5%

10%

15%

20%

# rating of users

RM

SE

imp

rove

men

t

SoMFHeteCFHeteMFDSR

[0,5] [6,20] [21,) All−5%

0

5%

10%

15%

20%

# rating of items

RM

SE

imp

rove

men

t

[0,5] [6,20] [21,) All−5%

0

5%

10%

15%

20%

# rating of users&items

RM

SE

imp

rove

men

t

(a) (b) (c)

(d) (e) (f)

Fig. 3 MAE&RMSE improvement against PMFonvarious cold-start levels.aUsers_MAE,b Items_MAE, cUsers&Items_MAE,dUsers_RMSE,e Items_RMSE, f Users&Items_RMSE

It is obvious that DSR with similarity constraint alwaysachieves better performances than the basic matrix fac-torization (i.e., PMF). When integrating more meta-paths,DSR will has better performances (see UU∗ and MM∗).Observing the experimental results of DSR with differentmeta-paths, we can find that the meaningful and dense sim-ilarity generated by meta-path will make DSR have betterperformances. For example, the UMTMU gets a slightlybetter result, while UGU gets the worst results. This is con-sistent with the real world, since users liking the same typeof movies (i.e., UMTMU) always have similar preferencesfor movies. However, users in the same group (i.e., UGU)together with various reasons may have huge differences inmovie preference. Another example is that theMTMpath hasbetter performance, while the MUUM path has the worst. Itsuggests that the similarity of movies with the same type(i.e., MTM) plays a leading role in movie recommendation,while the similarity of movies generated by the same movieswatched by two friends is not such meaningful and useful.

6.6 Parameter study on α, β and γ

TheDSRmodel is based on the low-rankmatrix factorizationframework, and the similar regularization on users and itemsis applied to constrain the model learning process. The rele-vant parameters of the basic matrix factorization have beenstudied in other matrix factorization methods. In this section,we study α and β which are the parameters of dual similarity

regularization on Douban data set. Considering the impactof γ in dual similarity regularization, we also take parameterstudy experiments on it.

Figure 5 shows that the impacts of α and β on MAE andRMSE are quite similar. When the values of α and β are botharound 10, the experiment has the best performance. Whenthe values of α and β are quite large or small, the resultsare not optimal. When α and β are set the proper values (inour experiments they are both 10), regularization and ratinginformation take effect on the learning process simultane-ously so that the experiments could get better performance.It indicates that integrating the similarity information ofusers and items in a HIN has a significant impact on rec-ommender systems. Compared with the optimal results, theexperimental results decline sharply when the values of α

and β are increased from 10. On the other hand, when α

and β are quite small, DSR performs like basic matrix fac-torization method but the experimental results are not toobad.

Figure 6 shows the value of γ and also has a significantinfluence on experimental result. γ controls the radial inten-sity of Gaussian function on dual similarity regularization,which further controls the similarity of latent factors of twousers (or items). Nineteen different values of γ are set (from0.1 to 10) in our experiments, and optimal values are set forother parameters. We can see that when the value of γ is in[0.5, 4], it has good performance. When the value of γ tendsto be quite large or small, the constraints on users or items

123

46 Int J Data Sci Anal (2017) 3:35–48

0.6

0.65

0.7

0.75

PMF

UMAMUULU

UGUUMU

UMDMU UU

UMTMUUU*

MAMMDM

MTMMUM

MUUMMM*

MA

E80%60%40%

0.75

0.8

0.85

0.9

0.95

PMF

UMAMUULU

UGUUMU

UMDMU UU

UMTMUUU*

MAMMDM

MTMMUM

MUUMMM*

RM

SE

80%60%40%

(a)

(b)Fig. 4 Impact of different meta-paths. a MAE, b RMSE

cannot work well, which leads to poor performance. Consid-ering the experimental results on MAE and RMSE, we setγ = 1 in the former experiments.

7 Conclusions

In this paper, we analyzed the limitations of social regulariza-tion, including three aspects: First, the social regularizationonly takes constraint on users with their social relation; sec-ond, much available information of users and items can beused in HIN, while the social regularization did not make

00.001

0.010.1

1 10100

1000

00.001

0.010.1

110

10010000.5

1

1.5

2

βα

MA

E

00.001

0.010.1

1 10100

1000

00.001

0.010.1

110

10010000.5

1

1.5

2

2.5

βα

RM

SE

(a)

(b)

Fig. 5 Parameter study on α and β. a MAE, b RMSE

0 0.2 0.4 0.6 0.8 1 2 4 6 8 100.55

0.60

0.65

0.70

0.75

0.80

0.85

0.90

0.95

γ

MA

E &

RM

SE

MAERMSE

Fig. 6 Parameter study on γ

full use of it; third, the social regularization cannot workwell for users with low similarity. Therefore, we design adual similarity regularization whose basic idea is to enforcethe constraints on both similar and dissimilar objects simul-

123

Int J Data Sci Anal (2017) 3:35–48 47

taneously. Then, encouraged by the successful of low-rankmatrix factorization method, we employ the similarity reg-ularization on the low-rank matrix factorization frameworkand proposed theDSRmethod, which integrates much usefulinformation and takes constraint on users more effectively.Experiments on three real-world data sets validate the effec-tiveness of DSR, especially on alleviating the cold-startproblem.

Acknowledgements This work is supported in part by NationalKey Basic Research and Department (973) Program of China (No.2013CB329606), the National Natural Science Foundation of China(Nos. 71231002, 61375058, 11571161, 11371115), the CCF-TencentOpen Fund, the Co-construction Project of BeijingMunicipal Commis-sion of Education, and Shenzhen Sci. Tech Fund No. JCYJ20140509143748226. The work of J. Li is supported by the NSF of China(No. 11571161) and Sci-Tech and Innov. Committee Shenzhen (No.JCYJ20160530184212170).

References

1. Balabanovic, M.: Content-based, collaborative recommendation.Commun. Acm 40(3), 66–72 (1997)

2. BellogíN, R., Cantador, I., Castells, P.: A comparative study ofheterogeneous item recommendations in social systems. Inf. Sci.221, 142–169 (2013)

3. Bernardi, L., Kamps, J., Kiseleva, J., Mller, M.J.: Solving the cold-start problem in recommender systems with social tags. Epl 92(2),28,002–28,007(6) (2010)

4. Burke, R., Vahedian, F., Mobasher, B.: Hybrid recommendationin heterogeneous networks. In: International Conference on UserModeling, Adaptation, and Personalization, pp. 49–60. Springer,New York (2014)

5. Cantador, I., Bellogin, A., Vallet, D.: Content-based recommenda-tion in social tagging systems. In: RecSys, pp. 237–240 (2010)

6. Cao, L., Yu, P.S.: Non-iid recommendation theories and systems.Intell. Syst. IEEE 31(2), 81–84 (2016)

7. Cplex, B.: Algorithms for non-negative matrix factorization. In:NIPS, pp. 556–562 (2010)

8. Diao, Q., Qiu, M., Wu, C.Y., Smola, A.J., Jiang, J., Wang, C.:Jointly modeling aspects, ratings and sentiments for movie recom-mendation (jmars). In: ACM SIGKDD International Conferenceon Knowledge Discovery and Data Mining, pp. 193–202 (2014)

9. Guo, Y.,Wang, X., Xu, C.: Lbmf: Log-bilinear matrix factorizationfor recommender systems. In: Pacific-Asia Conference on Knowl-edge Discovery and Data Mining, pp. 502–513 (2016)

10. Herlocker, J.L., Konstan, J.A., Borchers, A., Riedl, J.: An algorith-mic framework for performing collaborative filtering. In: SIGIR’99: Proceedings of the International ACM SIGIR Conference onResearch and Development in Information Retrieval, August 15–19, 1999, Berkeley, Ca, Usa, pp. 230–237 (1999)

11. Hong, L., Doumith, A.S., Davison, B.D.: Co-factorizationmachines: modeling user interests and predicting individual deci-sions in twitter. In: ACM International Conference on Web Searchand Data Mining, pp. 557–566 (2013)

12. Jaiwei, H.: Mining heterogeneous information networks: the nextfrontier. In: SIGKDD, pp. 2–3 (2012)

13. Jamali, M., Lakshmanan, L.V.: Heteromf: recommendation in het-erogeneous information networks using context dependent factormodels. In: WWW, pp. 643–653 (2013)

14. Lao, N., Cohen, W.W.: Fast query execution for retrieval modelsbased on path-constrained randomwalks. In: ACMSIGKDD Inter-

national Conference on Knowledge Discovery and Data Mining,pp. 881–888 (2010)

15. Leroy, V., Cambazoglu, B.B., Bonchi, F.: Cold start link prediction.In: Proceedings of the 16th ACM SIGKDD international confer-ence onKnowledge discovery and datamining, pp. 393–402. ACM(2010)

16. Li, F., Xu, G., Cao, L.: Coupled matrix factorization within non-iidcontext. In: Pacific-Asia Conference on Knowledge Discovery andData Mining, pp. 707–719. Springer, New York (2015)

17. Luo, C., Pang,W.,Wang, Z.: Hete-cf: social-based collaborative fil-tering recommendation using heterogeneous relations. In: ICDM,pp. 917–922 (2014)

18. Ma, H., King, I., Lyu, M.R.: Learning to recommend with socialtrust ensemble. In: SIGIR, pp. 203–210 (2011)

19. Ma, H., Yang, H., Lyu, M.R., King, I.: Sorec: Social recom-mendation using probabilistic matrix factorization. In: CIKM, pp.931–940 (2008)

20. Ma, H., Zhou, D., Liu, C., Lyu, M.R., King, I.: Recommendersystems with social regularization. In: WSDM, pp. 287–296(2011)

21. Salakhutdinov, R., Mnih, A.: Probabilistic matrix factorization. In:NIPS, pp. 1257–1264 (2012)

22. Sarwar, B., Karypis, G., Konstan, J., Riedl, J.: Item-based col-laborative filtering recommendation algorithms. In: InternationalConference on World Wide Web, pp. 285–295 (2001)

23. Schein, A.I., Popescul, A., Ungar, L.H., Pennock, D.M.: Methodsand metrics for cold-start recommendations. In: SIGIR 2002: Pro-ceedings of the International ACMSIGIR Conference on Researchand Development in Information Retrieval, August 11–15, 2002,Tampere, Finland, pp. 253–260 (2002)

24. Shi, C., Kong, X., Huang, Y., Yu, P.S., Wu, B.: Hetesim: A generalframework for relevance measure in heterogeneous networks. In:IEEETransactions onKnowledge andData Engineering, pp. 2479–2492 (2014)

25. Shi, C., Li, Y., Zhang, J., Sun, Y., Yu, P.S.: A survey of heteroge-neous information network analysis. Comput. Sci. 134(12), 87–99(2015)

26. Shi, C., Liu, J., Zhuang, F., Philip, S.Y., Wu, B.: Integrating het-erogeneous information via flexible regularization framework forrecommendation. In: Knowledge and Information Systems, pp. 1–25 (2016)

27. Shi, C., Zhang, Z., Luo, P., Yu, P.S., Yue, Y., Wu, B.: Semantic pathbased personalized recommendation on weighted heterogeneousinformation networks. In: The ACM International, pp. 453–462(2015)

28. Shi, C., Zhou, C., Kong, X., Yu, P.S., Liu, G., Wang, B.: Hetere-com: a semantic-based recommendation system in heterogeneousnetworks. In: Proceedings of the 18th ACMSIGKDD InternationalConference on Knowledge Discovery and Data Mining, pp. 1552–1555. ACM, New York (2012)

29. Srebro, N., Jaakkola, T.: Weighted low-rank approximations. In:ICML, pp. 720–727 (2003)

30. Sun, Y., Han, J., Yan, X., Yu, P., Wu, T.: Pathsim: meta path-basedtop-k similarity search in heterogeneous information networks. In:VLDB, pp. 992–1003 (2011)

31. Wang, C., Cao, L., Wang, M., Li, J., Wei, W., Ou, Y.: Couplednominal similarity in unsupervised learning. In: Proceedings of the20th ACM International Conference on Information and Knowl-edge Management, pp. 973–978. ACM, New York (2011)

32. Wu,Y.,Dubois,C., Zheng,A.X., Ester,M.:Collaborative denoisingauto-encoders for top-n recommender systems. In: ACM Interna-tional Conference on Web Search and Data Mining (2016)

33. Yang, C., Zhou, Y., Chen, L., Zhang, X., Chiu, D.M.: Social groupbased video recommendation addressing the cold-start problem.In: Pacific-Asia Conference on Knowledge Discovery and DataMining, pp. 515–527 (2016)

123

48 Int J Data Sci Anal (2017) 3:35–48

34. Yang, X., Steck, H., Liu, Y.: Circle-based recommendation inonline social networks. In: KDD, pp. 1267–1275 (2012)

35. Yu, X., Ren, X., Gu, Q., Sun, Y., Han, J.: Collaborative filteringwith entity similarity regularization in heterogeneous informationnetworks. In: IJCAI-HINA Workshop (2013)

36. Yu, X., Ren, X., Sun, Y., Gu, Q., Sturt, B., Khandelwal, U., Norick,B., Han, J.: Personalized entity recommendation: a heterogeneousinformation network approach. In: ACM International Conferenceon Web Search and Data Mining, pp. 283–292 (2014)

37. Yu, X., Ren, X., Sun, Y., Sturt, B., Khandelwal, U., Gu, Q.,Norick, B., Han, J.: Recommendation in heterogeneous informa-tion networks with implicit user feedback. In: ACMConference onRecommender Systems, pp. 347–350 (2013)

38. Zhang, D., Hsu, C.H., Chen, M., Chen, Q., Xiong, N., Lloret,J.: Cold-start recommendation using bi-clustering and fusion forlarge-scale social recommender systems. IEEE Trans. Emerg. Top.Comput. 2(2), 239–250 (2014)

39. Zhang, J., Tang, J., Liang,B.,Yang,Z.,Wang, S., Zuo, J., Li, J.: Rec-ommendation over a heterogeneous social network. In: The NinthInternational Conference on Web-Age Information Management,2008. WAIM’08, pp. 309–316. IEEE (2008)

40. Zheng, J., Liu, J., Shi, C., Zhuang, F., Li, J., Wu, B.: Dual similarityregularization for recommendation. In: Pacific-Asia Conference onKnowledge Discovery and Data Mining, pp. 542–554. Springer,New York (2016)

123