mohsen jamali, martin ester simon fraser university vancouver, canada acm recsys 2010

A Matrix Factorization Technique with Trust Propagation for Recommendation in Social NetworksMohsen Jamali, Martin EsterSimon Fraser UniversityVancouver, Canada

ACM RecSys 2010

Outline

Introduction Matrix Factorization Models

Basic MF Model Social Trust Ensemble Model

The SocialMF Model Data Sets Experiments Conclusions

2Mohsen Jamali, Social Matrix Factorization

Introduction

Need For Recommenders Input Data

A set of users U={u1, …, uN}

A set of items I={i1, …, iM}

The rating matrix R=[ru,i]NxM

Problem Definition: Given user u and target item i Predict the rating ru,i

Collaborative Filtering Approach

3Mohsen Jamali, Social Matrix Factorization

Introduction (cont.)

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 Factorization 4

Recommendation in Social Networks

Memory based approaches for recommendation in social networks [Golbeck, 2005] [Massa et.al.

2007] [Jamali et.al.

2009] [Ziegler, 2005]

Mohsen Jamali, Social Matrix Factorization 5

A Sample Social Rating Network

Matrix Factorization

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 Factorization 6

U and V have normal priors

Social Trust Ensemble [2009]

Mohsen Jamali, Social Matrix Factorization 7

Social Trust Ensemble (cont.)

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 Factorization 8

The SocialMF Model

Social Influence behavior of a user u is affected by his direct neighbors Nu.

Latent characteristics of a user depend on his neighbors.

Tu,v is the normalized trust value.

Mohsen Jamali, Social Matrix Factorization 9

The SocialMF Model (cont.)

Mohsen Jamali, Social Matrix Factorization 10

The SocialMF Model (cont.)

Mohsen Jamali, Social Matrix Factorization 11

The SocialMF Model (cont.)

Mohsen Jamali, Social Matrix Factorization 12

The SocialMF Model (cont.)

Mohsen Jamali, Social Matrix Factorization 13

The SocialMF Model (cont.)

Mohsen Jamali, Social Matrix Factorization 14

The SocialMF Model (cont.)

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 Factorization 15

Data Sets

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 Factorization 16

Data Sets (cont.)

Mohsen Jamali, Social Matrix Factorization 17

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

Experimental Setups

5-fold cross validation Using RMSE for evaluation Comparison Partners

Basic MF STE CF

Model parameters SocialMF: STE:

Mohsen Jamali, Social Matrix Factorization 18

Results for Epinions

Gain over STE: 6.2%. for K=5 and 5.7% for K=10

Mohsen Jamali, Social Matrix Factorization 19

CF BaseMF STE SocialMF1

1.021.041.061.08

1.11.121.141.161.18

1.2

k=5k=10R

MS

E

Results for Flixster

SocialMF gain over STE (5%) is 3 times the STE gain over BasicMF (1.5%)

Mohsen Jamali, Social Matrix Factorization 20

CF BaseMF STE SocialMF0.76

0.78

0.8

0.82

0.84

0.86

0.88

0.9

0.92

k=5k=10R

MS

E

Sensitivity Analysis on λT

Mohsen Jamali, Social Matrix Factorization 21

Sensitivity Analysis for Epinions

Sensitivity Analysis on λT

Mohsen Jamali, Social Matrix Factorization 22

Sensitivity Analysis for Flixster

Experiments on Cold Start Users

Mohsen Jamali, Social Matrix Factorization 23

RMSE values on cold start users (K=5)

CF BaseMF STE SocialMF1

1.1

1.2

1.3

1.4

Epinions

Experiments on Cold Start Users

Mohsen Jamali, Social Matrix Factorization 24

RMSE values on cold start users (K=5)

CF BaseMF STE SocialMF0.95

11.051.1

1.151.2

1.25

Flixster

Experiments on Cold Start Users

Mohsen Jamali, Social Matrix Factorization 25

Flixster Epinions

Cold Start Users 0.085 0.115

All Users 0.05 0.062

-1.00%

1.00%

3.00%

5.00%

7.00%

9.00%

11.00%

RMSE Gain of SocialMF over STE

Analysis of Learning Runtime

SocialMF: STE: SocialMF is faster by factor

Mohsen Jamali, Social Matrix Factorization 26

N # of Users

K Latent Feature Size

Avg. ratings per user

Avg. neighbors per user

rt

Conclusion

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 Factorization 27

Mohsen Jamali. Using Trust Networks to Improve Top-N Recommendation

28

Thank you!