dm587 collaborative filtering with social local models · collaborative filtering with social local...
TRANSCRIPT
CollaborativeFilteringwithSocialLocalModels
Huan Zhao
Joint work with Quanming Yao, JamesT.KwokandDik Lun LeeDepartmentofCSE,HKUST,HongKong
1
Recommender System
2
Products Recommendation
Recommender System
3
Questions and AnswersRecommendation
Recommender System
4
RestaurantsRecommendation
Recommender System
• Recommendersystems (RS) areeverywhere.• They are not only useful for people, but alsocreate huge revenues for companies.
• The most popular RS method is collaborativefiltering (CF).– User-based CF– Item-based CF–Matrix Factorization (MF)
5
MatrixFactorization (MF)• The assumption is that users’ preferences arecontrolled by a small number of factors.
6https://buildingrecommenders.wordpress.com/2015/11/18/overview-of-recommender-algorithms-part-2/
min$,&
12))𝐼+, 𝑅+, − 𝑢+𝑏,
1 +4
,56
7
+56
𝜆62 𝑈 :
1 +𝜆12 𝐵 :
1
Problems of MF• Sparsity of the rating matrix.–More than 99% entries are missing.
• Cold Start.– Some users or items have no ratings.
• More importantly, the low-rank assumption maybe too restrictive.– Lee et.al. relaxed this assumption to propose the locallow rank matrix approximation (LLORMA) framework.[Lee et.al. ICML 2013]
7
LLORMA• The rating matrix is composed of a number ofsmaller matrices which are of low rank.
• The rating is then predicted by weighted ensembleof predictions from all submatrices.
8
Problems of LLORMA• Meaningless of submatrices– Randomly select anchor points from the rating matrix.– Select other points within some distance threshold.
• Inconsistence of submatrices.
• Space and computational cost.– Need to save pairwise similarities of all points in allsubmatrices.
9
Our Work (Social Local Models)• We are the first to integrate social connectionswith LLORMA.– It can enjoy the advantages of both socialrecommendation and local low rank assumption.
• Meaningfulness of submatrices.– Social communities can explain the local models.
• Social regularization with LLORMA can furtherimprove the recommending performance.
10
Our Work (Social Local Models)
11
• Right part is LLORMA.• Left part motivates our framework.
Framework• Identifysocialgroupsfromthesocialgraphandthenconstructsubmatricesbasedonthegroups.
• ApplyMFtoallthesubmatrices,independently,andobtain multiplegroupsofuser-specificanditem- specificlatentfeatures.
• Predictthemissingratingsusingtheensembleofpredictionsfromallsubmatrices.
12
Submatrices Construction• Heuristic methods.– Buildthesocialgroupsbasedonthefactthatusers’influencestoeachothercanpropagatethroughthenetworks.
– Submatrices are constructed by firstly select influentialusers, called connectors, and their friends within a fixednumber of hops.
• Different methods to select connectors.– Hub: those with the most friends.– Random: randomly selection.– Random-Hub: randomly selection from a set of Hub users.– Greedy: eachtimeweselectaconnector,weselectfromthosenotyetcoveredbyexisting connectors.
13
Submatrices Construction• Systematic methods.– Overlapping community detections methods can beexploited.
– BIGCLAM are utilized for its scalability and goodperformance.[Yang et.al. ICDM 2013]
• After we create social groups, for each group, weselect the users and the items they rate toconstruct a submatrix.
14
Submatrices Factorization• Social LOcal Matrix Approximation (SLOMA).– Apply MF to each submatrix independently.
• SLOMA++.– Add social regularization to each submatrixfactorization.
15
Experimental Results• Datasets.
• Evaluation metrics.– The smaller, the better.
16
Experimental Results
• Our proposed methods outperform baselines consistently,which demonstrates the effectiveness of our framework.
• Comparing SLOMA and LLORAM, the performance gain isobtained by incorporating the social connections.
• Comparing SLOMA++ and SocReg, the performance gain isobtained by the local low rank assumption.
17
Experimental Results
18
• Impact of number of local models.–When the number is smaller than 10, the performanceis weak.
–When the number is larger, e.g., 50, the performance isconsistently good.
Experimental Results
19
• Impact of number of hops.–When the number is smaller than 3, the performanceis weak.
–When the number is larger than 3, the performance isconsistently good.
Experimental Results• Impact of Connector Selection methods.– Hub and Greedy are the best two methods.– Community-based one is the worst.
20
Conclusion• We propose social local models to enhanceLLORMA.
• Social recommendation and local low rankassumption can both benefit the recommendingperformance.
• Experimental results have demonstrates theeffectiveness of our SLOMA and SLOMA++.
21