contextual evsm: a context-aware content-based recommendation framework based on distributional...

Post on 15-Jul-2015

1.257 Views

Category:

Technology

0 Downloads

Preview:

Click to see full reader

TRANSCRIPT

EC-WEB 2013 - 14th International Conference on Electronic Commerce and Web Technologies Prague (Czech Republich) - 28.08.13

Contextual eVSM: a context-aware content-based recommendation framework based on

distributional semantics

Cataldo Musto, Giovanni Semeraro, Pasquale Lops, Marco de Gemmis

scenario.

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

Prague, EC-Web conferenceC. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

dinner timeC. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

me and Pasquale like sushi.C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

what does TripAdvisor suggest?

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

Good PriceC. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

Good ReviewsC. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

....Strasbourg?C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

612 km far away.

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

612 km far away.Too much :-)

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

we ate pizza, as usual.C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

why did the TripAdvisor recommendation engine fail?

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

it doesn’t take into account contextual information.

context plays a key role in several recommendation tasks.

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

I attended last Sigur Ròs concert in Rome, so I like them

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

I attended last Sigur Ròs concert in Rome, so I like them

but their ambient music is not good if I need music recommendation for my running session!

a real-world recommendation engine needs to take into account

contextual information

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

what is context?

• Over 150 definitions, spread across several domains (Bazire and Brezillon, 2005)

• Recommender Systems area

• “A set of factors that have influence on user perception and acceptance of a particular item”

• A fixed set of dimensions with appropriate attributes

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

what is context?

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

task

what is context?

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

task

what is context?

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

task company

what is context?

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

task company

what is context?

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

task moodcompany

what is context?

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

task company mood

a real-world recommendation engine needs to take into account

contextual information

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

Our contribution

contextual eVSMa context-aware content-based recommendation

framework based on distributional semantics

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

timeline

2013: contextual eVSM

2010-2012: eVSM(*)a content-based recommendation framework

based on distributional semantics

context-aware extension of eVSM

(*) Cataldo Musto: Enhanced vector space models for content-based recommender systems. RecSys 2010: 361-364

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

eVSMa brief recap

eVSM

• Content-based Recommendation Framework

• (semantic) vector space representation based on distributional models

• negative information modeled through quantum negation operator

• recommendation seen as a form of similarity in vector spaces

• four different profiling models

• cosine similarity to get the top-k recommendations

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

cornerstones

distributional models(Firth, 1957)

Firth, J.R. A synopsis of linguistic theory 1930-1955. In Studies in Linguistic Analysis, pp. 1-32, 1957.

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

cornerstone 1

“meaning is its use”

L.Wittgenstein(Austrian philosopher)

distributional semantics

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

insightby analyzing large corpus of textual data it is possible to infer information about the usage (about the meaning) of the terms.example

distributional models

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

distributional hypothesis

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

“terms that occur in similar contexts share a similar meaning”

distributional models

c1 c2 c3 c4 c5 c6

rock ✔ ✔ ✔

post rock ✔ ✔

jazz ✔

classical ✔ ✔ ✔

term/context matrix (WordSpace)

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

distributional modelsrock vs. post rock = good overlap

c1 c2 c3 c4 c5 c6

rock ✔ ✔ ✔

post rock ✔ ✔

jazz ✔

classical ✔ ✔ ✔

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

distributional modelsrock vs. classical = no overlap

c1 c2 c3 c4 c5 c6

rock ✔ ✔ ✔

post rock ✔ ✔

jazz ✔

classical ✔ ✔ ✔

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

• Key: definition of what is the ‘context’

• Different granularities are possible

• Document

• Paragraph

• Sentence

• Sliding window of words

distributional models

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

rockalternative-rock

post-rock

classical

exampleWordSpace

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

representation of documents (*) can be inferred by combining the representation of the terms (**) occurring in the document.

(*) documents = artists(**) terms = tags

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

distributional models

c1 c2 c3 c4 c5 c6 c7 c8 c9

t2 ✔ ✔ ✔ ✔

t3 ✔ ✔ ✔

d1 ✔ ✔ ✔ ✔ ✔

term/context matrix (DocSpace)

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

Coldplay

RadioheadKings of Leon

Lady Gaga

exampleDocSpace

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

distributional models

• Features

• semantic vector space representation of terms and documents (user profiles and items !)

• light semantics, based on term co-occurrences in large corpus of data

• based on distributional hypothesis

• totally unsupervised

• just based on the analysis of term distribution

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

quantum negation(Widdows, 2007)

cornerstone 2

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

negation in VSMs

•Widdows proposed a different point of view

• Negation view as a form of orthogonality between vectors

• Vision inherited from Quantum Logic

state of the art

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

negation in VSMs

• Some theory

• Given vector a and vector b• Through quantum negation it is possible to define a

vector a not b (a ∧¬b)

• Projection of vector a on the subspace

orthogonal to those generated by vector b

Quantum Negation

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

negation in VSMs

• Two terms can be considered as mutually unrelevant if they never co-occur

• Two documents can be considered as mutually unrelevant if they have no features in common

• No features in common ⇒ scalar product = 0

•Orthogonality

Widdows’ insights

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

quantum negation

• Vector A models positive feedbacks

• Information about what a user likes

• Vector B models negative feedbacks

• Information about what a user does not like

• Vector A not B combines both information sources

application to CBRS

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

recommendation step

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

cornerstone 3

ThresholdRatingItems

Higher weight given to the documents with higher rating

User ProfilesWeighted Random Indexing-based (w-RI)

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

VSM representation of wQN-based profile for user u

Positive User Profile Vector

Negative User Profile Vector

User ProfilesWeighted Quantum Negation-based (w-QN)

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

scenario

football news

sports news

politics news

politics news

user profile

Recommendation task seen as similarity

calculation between vectors

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

scenario

football news

sports news

politics news

politics news

user profile

football and sports news are recommended to

the target user

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

size=400 - Movielens dataset

Gap always around 1%

84

84,75

85,5

86,25

87

p@1 P@3 P@5 P@10

84,7584,7

84,97

85,39

84,5884,7

84,85

85,27

84,5

84,77

84,4384,47

85,5285,58

86,0185,94eVSM VSMLSI Bayes

experiment(*)

Cataldo Musto - Enhanced Vector Space Models for Content-based Recommender Systems - Ph.D. defense - University of Bari Aldo Moro, Italy - 08.06.12

(*) Cataldo Musto. Enhanced Vector Space Models for content-based Recommender Systems. Ph.D dissertation, 2012

84

84,75

85,5

86,25

87

p@1 P@3 P@5 P@10

84,7584,7

84,97

85,39

84,5884,7

84,85

85,27

84,5

84,77

84,4384,47

85,5285,58

86,0185,94eVSM VSMLSI Bayes

size=400 - Movielens dataset

Significant Improvement

Cataldo Musto - Enhanced Vector Space Models for Content-based Recommender Systems - Ph.D. defense - University of Bari Aldo Moro, Italy - 08.06.12

experiment(*) (*) Cataldo Musto. Enhanced Vector Space Models for content-based Recommender Systems. Ph.D dissertation, 2012

eVSM outperforms other state-of-the-art approaches

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

eVSM outperforms other state-of-the-art approaches

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

but it doesn’t take into account contextual information.

contextual eVSM

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

how to make the eVSM context-aware?

context-aware RSs

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

state of the art: pre and post-filtering

we implemented two contextualization strategies

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

both pre and post-filtering

context-aware eVSM

•Microprofiling

•based on contextual pre-filtering

• Insight: to filter data before building user profiles, building a separate user profile for each contextual dimension

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

context-aware eVSM

•Contextual eVSM

•based on contextual post-filtering

•Insight: to calculate un-contextualized recommendation and to re-rank them according to contextual constraints

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

if a user needs suggestions for a restaurant for a romantic dinner, only her ratings expressed for previous romantic dinners have to be taken

into account

micro-profilinginsight

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

micro-profilingcontext-aware eVSM

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

micro-profilingcontext-aware eVSM

user

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

micro-profilingcontext-aware eVSM

contextual dimension (e.g task)and value (e.g. running)

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

micro-profilingcontext-aware eVSM

sum over all the items rated under that specific contextual constraints

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

micro-profilingcontext-aware eVSM

weighted with the (normalized) rating provided by the user

contextual eVSM insight

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

context is just a factor that can (positively or negatively) influence user preference on a certain item

contextual eVSM insight

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

context is just a factor that can (positively or negatively) influence user preference on a certain item

context-aware recommendation can be obtained by combining an uncontextual user profile with a

vector space representation of the target context

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

contextual eVSMcontext-aware eVSM

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

contextual eVSMcontext-aware eVSM

user profile is the combination of two components

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

contextual eVSMcontext-aware eVSM

(1) uncontextual user profile

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

contextual eVSMcontext-aware eVSM

(2) vector space representation of the context

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

contextual eVSMcontext-aware eVSM

weight of each component

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

contextual eVSMcontext-aware eVSM

if alpha = 1, we have uncontextual recommendation since context weight = 0 !

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

contextual eVSMcontext-aware eVSM

problem: how to build it ?

• For each contextual dimension, it exists a set of terms that is descriptive of items relevant in that specific context

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

context representationassumptions

• For each contextual dimension, it exists a set of terms that is descriptive of items relevant in that specific context

• e.g. candlelight or seaview are relevant features if I’m looking for a restaurant for a romantic night !

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

context representationassumptions

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

context representationidea

to represent the context as the combination of the terms occurring in the descriptions of the items labeled as relevant under that specific contextual situation

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

context representationformula

PRE-WRI is exactly the combination of the terms occurring in the descriptions of the items labeled as relevant under that specific contextual situation !

we adopted PRE-WRI as vector space representation of the context

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

rationalecontextual eVSM

if the user didn’t express any preference in that specific context, the vector is null, so she will receive

uncontextual recommendation

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

rationalecontextual eVSM

if the user didn’t express any preference in that specific context, the vector is null, so she will receive

uncontextual recommendationotherwise, a greater weight will be given to the

features that are informative and relevant in the target context, that is to say, those contained in the

context vector!

experimental evaluation.

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

• Experiments

• (1) Does contextual eVSM outperform eVSM ?

• (2) How does our approach performs with respect to current literature?

experimental design

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

•Movie recommendation (IMDB data)

• 202 movies (textual features crawled from Wikipedia)

• 62 users and 1457 ratings

• 4 contextual dimensions

• TIME (weekend, weekday)

• PLACE (theather, home)

• COMPANION (alone, friends, boyfriend, family)

• MOVIE-RELATED (release week or not)

experimental designdataset

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

• Dataset and experimental settings replicate Adomavicius’ experiment (*)

• Evaluation over 9 different contextual settings

• Home, Friends, Non-release, Weekend, Weekday, GBFriends, TheatherWeekend and TheatherFriends

• Experimental protocol: bootstrapping

• 29/30th of the data as training

• 1/30th as test

• Randomly generated, 500 runs

experimental designsettings

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

(*) G.Adomavicius et al. , Incorporating contextual

information in recommender systems using a multi-

dimensional approach. ACM Trans. Inf. Systems, 2005

• eVSM settings

• Uncontextual baselines

• WRI and WQN

• Pre-filtering

• PRE-WRI and PRE-WQN

• Post-filtering:

• CONTEXT-WRI and CONTEXT-WQN

• alpha=0.5, alpha=0.8

• 8 settings for each run

experimental designparameters

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

experiment 1selection of results - HOME segment

WRI

PRE-WRI

CONT-WRI-0.5

CONT-WRI-0.8

WQN

PRE-WQN

CONT-WQN-0.5

CONT-WQN-0.8

45 48,75 52,5 56,25 60

pre-filtering outperforms the baseline (it’s an exception, in the other segments it doesn’t !)

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

experiment 1selection of results - HOME segment

WRI

PRE-WRI

CONT-WRI-0.5

CONT-WRI-0.8

WQN

PRE-WQN

CONT-WQN-0.5

CONT-WQN-0.8

45 48,75 52,5 56,25 60

contextual eVSM improves the F1 measureC. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

experiment 1selection of results - HOME segment

WRI

PRE-WRI

CONT-WRI-0.5

CONT-WRI-0.8

WQN

PRE-WQN

CONT-WQN-0.5

CONT-WQN-0.8

45 48,75 52,5 56,25 60

alpha=0.8 better than alpha 0.5C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

experiment 1selection of results - HOME segment

WRI

PRE-WRI

CONT-WRI-0.5

CONT-WRI-0.8

WQN

PRE-WQN

CONT-WQN-0.5

CONT-WQN-0.8

45 48,75 52,5 56,25 60

contextual eVSM with negation is the best settingC. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

experiment 1selection of results - FRIEND segment

WRI

PRE-WRI

CONT-WRI-0.5

CONT-WRI-0.8

WQN

PRE-WQN

CONT-WQN-0.5

CONT-WQN-0.8

45 47,5 50 52,5 55

similar outcomes

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

experiment 1selection of results - FRIEND segment

WRI

PRE-WRI

CONT-WRI-0.5

CONT-WRI-0.8

WQN

PRE-WQN

CONT-WQN-0.5

CONT-WQN-0.8

45 47,5 50 52,5 55

pre-filtering doesn’t improve the accuracyC. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

experiment 1selection of results - FRIEND segment

WRI

PRE-WRI

CONT-WRI-0.5

CONT-WRI-0.8

WQN

PRE-WQN

CONT-WQN-0.5

CONT-WQN-0.8

45 47,5 50 52,5 55

contextual eVSM generally does

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

experiment 1selection of results - FRIEND segment

WRI

PRE-WRI

CONT-WRI-0.5

CONT-WRI-0.8

WQN

PRE-WQN

CONT-WQN-0.5

CONT-WQN-0.8

45 47,5 50 52,5 55

contextual eVSM with negation and alpha=0.8 is always the best setting

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

experiment 1selection of results - NON RELEASE segment

WRI

PRE-WRI

CONT-WRI-0.5

CONT-WRI-0.8

WQN

PRE-WQN

CONT-WQN-0.5

CONT-WQN-0.8

45 48 51 54 57

contextual eVSM with negation and alpha=0.8 is always the best setting

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

experiment 1selection of results - NON RELEASE segment

WRI

PRE-WRI

CONT-WRI-0.5

CONT-WRI-0.8

WQN

PRE-WQN

CONT-WQN-0.5

CONT-WQN-0.8

45 48 51 54 57

typically, alpha=0.8 is better than alpha 0.5

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

experiment 1selection of results - NON RELEASE segment

WRI

PRE-WRI

CONT-WRI-0.5

CONT-WRI-0.8

WQN

PRE-WQN

CONT-WQN-0.5

CONT-WQN-0.8

45 48 51 54 57

outcome: context has just a little influence on user perception and acceptance. Uncontextual preferences are still the “core”.

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

experiment 1selection of results - NON RELEASE segment

WRI

PRE-WRI

CONT-WRI-0.5

CONT-WRI-0.8

WQN

PRE-WQN

CONT-WQN-0.5

CONT-WQN-0.8

45 48 51 54 57

outcome: context has to be taken into account, but just a little!

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

experiment 1selection of results - THEATHER segment

WRI

PRE-WRI

CONT-WRI-0.5

CONT-WRI-0.8

WQN

PRE-WQN

CONT-WQN-0.5

CONT-WQN-0.8

45 47,25 49,5 51,75 54

However, in some setting contextual eVSM without negation is the best

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

experiment 1selection of results - THEATHER segment

WRI

PRE-WRI

CONT-WRI-0.5

CONT-WRI-0.8

WQN

PRE-WQN

CONT-WQN-0.5

CONT-WQN-0.8

45 47,25 49,5 51,75 54

experiments showed a clear relationship between the amount of negative ratings and the best configurations

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

experiment 1selection of results - THEATHER segment

WRI

PRE-WRI

CONT-WRI-0.5

CONT-WRI-0.8

WQN

PRE-WQN

CONT-WQN-0.5

CONT-WQN-0.8

45 47,25 49,5 51,75 54

when the dataset is well balanced, cont-WQN outperforms cont-WRI, otherwise when few negative data are available, cont-WRI is the best

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

experiment 1selection of results - THEATHER+FRIENDS segment

WRI

PRE-WRI

CONT-WRI-0.5

CONT-WRI-0.8

WQN

PRE-WQN

CONT-WQN-0.5

CONT-WQN-0.8

40 41,75 43,5 45,25 47

contextual eVSM didn’t improve F1 measure in just 1 contextual segment out of 9

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

experiment 1 - outcomecontextual eVSM improves the predictive accuracy of eVSM

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

experiment 2comparison with state-of-the-art

HOME

FRIENDS

WEEKEND

THEATHER

NONRELEASE

WEEKDAY

GBFRIEND

THEAT-WEEK

THEAT-FRIENDS

37 45,25 53,5 61,75 70

comparison with Adomavicius’ approach based on CFC. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

experiment 2comparison with state-of-the-art

HOME

FRIENDS

WEEKEND

THEATHER

NONRELEASE

WEEKDAY

GBFRIEND

THEAT-WEEK

THEAT-FRIENDS

37 45,25 53,5 61,75 70

eVSM outperforms CF in 6 segments out of 9

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

experiment 2 - outcomecontextual eVSM improves the predictive accuracy

of state-of-the-art approaches based on CF

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

recap.

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

recap• context-aware eVSM

• baseline: eVSM

• content-based recommendation framework

• distributional semantics + quantum negation

• evolution: contextual eVSM

• comparison of two approaches for context-aware eVSM

• exact pre-filtering and weighted post-filtering

• large experimental evaluation, comparison with state of the art approaches

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

contextual eVSM overcomes state-of-the-art approaches

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

future research.C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

evaluation with different datasets

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

open knowledge sources and linked data for CBRS.

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

evaluation with user-based metrics(serendipity, novelty, unexpectedness)

C. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

questions?Cataldo Musto, Ph.D

cataldo.musto@uniba.itC. Musto, G. Semeraro, P. Lops, M. de Gemmis. - Contextual eVSM: a context-aware content-based recommendation framework based on distributional semantics - EC WEB 2013 - Prague, Czech Republic - 28.08.2013

top related