1. Mohsen Jamali, Martin Ester Simon Fraser University Vancouver, Canada ACM RecSys 2010

2.  Introduction  Matrix Factorization Models  Basic MF Model  Social Trust Ensemble Model  The SocialMF Model  Data Sets  Experiments  Conclusions 2Mohsen Jamali, Social Matrix Factorization

3.  Need For Recommenders  Input Data  A set of users U={u 1, …, u N }  A set of items I={i 1, …, i M }  The rating matrix R=[r u,i ] NxM  Problem Definition:  Given user u and target item i  Predict the rating r u,i  Collaborative Filtering Approach 3Mohsen Jamali, Social Matrix Factorization

4.  Social Networks Emerged Recently  Independent source of information  Motivation of SN-based RS  Social Influence: users adopt the behavior of their friends  Social Rating Network  Social Network  Trust Network Mohsen Jamali, Social Matrix Factorization4

5.  Memory based approaches for recommendation in social networks  [Golbeck, 2005]  [Massa et.al. 2007]  [Jamali et.al. 2009]  [Ziegler, 2005] Mohsen Jamali, Social Matrix Factorization5 A Sample Social Rating Network

6.  Model based approach  Latent features for users  Latent features for items Ratings are scaled to [0,1] g is logistic function Mohsen Jamali, Social Matrix Factorization6 U and V have normal priors

7. Mohsen Jamali, Social Matrix Factorization7

8.  Issues with STE  Feature vectors of neighbors should influence the feature vector of u not his ratings  STE does not handle trust propagation  Learning is based on observed ratings only. Mohsen Jamali, Social Matrix Factorization8

9.  Social Influence  behavior of a user u is affected by his direct neighbors N u.  Latent characteristics of a user depend on his neighbors.  T u,v is the normalized trust value. Mohsen Jamali, Social Matrix Factorization9

10. 10

11. Mohsen Jamali, Social Matrix Factorization11

12. Mohsen Jamali, Social Matrix Factorization12

13. Mohsen Jamali, Social Matrix Factorization13

14. Mohsen Jamali, Social Matrix Factorization14

15.  Properties of SocialMF  Trust Propagation  User latent feature learning possible with existence of the social network ▪ No need to fully observed rating for learning ▪ Appropriate for cold start users Mohsen Jamali, Social Matrix Factorization15

16.  Epinions – public domain  Flixster  Flixster.com is a social networking service for movie rating  The crawled data set includes data from Nov 2005 – Nov 2009  Available at http://www.cs.sfu.ca/~sja25/personal/datasets/ Mohsen Jamali, Social Matrix Factorization16

17. Mohsen Jamali, Social Matrix Factorization17  General Statistics of Flixster and Epinions  Flixster: 1M users, 47K items  150K users with at least one rating  Items: movies  53% cold start  Epinions: 71K users, 108K items  Items: DVD Players, Printers, Books, Cameras,…  51% cold start

18.  5-fold cross validation  Using RMSE for evaluation  Comparison Partners  Basic MF  STE  CF  Model parameters  SocialMF:  STE: Mohsen Jamali, Social Matrix Factorization18

19.  Gain over STE: 6.2%. for K=5 and 5.7% for K=10 Mohsen Jamali, Social Matrix Factorization19

20.  SocialMF gain over STE (5%) is 3 times the STE gain over BasicMF (1.5%) Mohsen Jamali, Social Matrix Factorization20

21. Mohsen Jamali, Social Matrix Factorization21 Sensitivity Analysis for Epinions

22. Mohsen Jamali, Social Matrix Factorization22 Sensitivity Analysis for Flixster

23. Mohsen Jamali, Social Matrix Factorization23 RMSE values on cold start users (K=5)

24. Mohsen Jamali, Social Matrix Factorization24 RMSE values on cold start users (K=5)

25. Mohsen Jamali, Social Matrix Factorization25

26.  SocialMF:  STE:  SocialMF is faster by factor Mohsen Jamali, Social Matrix Factorization26 N# of Users KLatent Feature Size Avg. ratings per user Avg. neighbors per user

27.  A model based approach for recommendation in social networks based on matrix factorization  Handling Trust Propagation  Appropriate for cold start users  Fast Learning phase  Improved quality for experiments on two real life data sets. Mohsen Jamali, Social Matrix Factorization27

28. Thank you! 28Mohsen Jamali. Using Trust Networks to Improve Top-N Recommendation