1. RECOMMENDER SYSTEMS WITH SOCIAL REGULARIZATION
HAO MA, DENGYONG ZHOU, CHAO LIU, MICHAEL R. LYU, IRWIN KING
MENGMENG YE
XIAO ZHANG
LU TIAN
TEXT MINING PAPER PRESENTATION

2. INTRODUCTION
If you want to find some good movies, music, or products. What will you do?

3. RECOMMENDER SYSTEM

4. TRADITIONAL RECOMMENDER SYSTEM
Assumption: Users are independent and have no relationship with each other
User-Item Rating Matrix

5. Ask Friends SOCIAL FRIENDSHIP
If you want to find some good movies, music, or products. What else will you do?
Ask Friends

6. TRUST-AWARE RECOMMENDER SYSTEM
Trust-aware Network
Unilateral relationship
Utilize trust information to enhance recommendation

7. SOCIAL RECOMMENDER SYSTEM
Social Network
Mutual relationship
Social relationship is more informative

8. LOW-RANK MATRIX FACTORIZATION
Goal: recover the missing ratings based on the observed ones
Next, I’m going to talk about the proposed method for social recommender system, which is the main focus of this paper. To begin with, let me briefly introduce how to deal with the traditional recommender system. So one of the most widely-used approach for recommendation is via low-rank matrix factorization. The idea is straightforward. It assumes the user-item matrix R is of low rank, which can be decomposed as product of a user feature matrix and a movie feature matrix. By doing so, if we can estimate the low-dimensional feature matrices U and V correctly, then we can predict any missing entries by the inner product of the corresponding user and movie features.
Accordingly, it can be formulated as the following optimization problem, where the first component measures the goodness of fit, that is how accurate the feature matrices can estimate the observed ratings, and the second component is a regularization term, similar to the regularization term for ridge regression, which can avoid overfitting. This optimization problem can be solved efficiently by gradient-based algorithm.
goodness of fit
avoid overfitting

9. SOCIAL REGULARIZATION 1
Average-based Regularization
Assume each user’s taste is close to the average taste of his/her friends
To incorporate the social friendship information, the authors proposed two social regularization methods.
Social Network
Sim(i,f) allows the user’s friends to have different tastes
It is insensitive to users whose friends have diverse tastes

10. SOCIAL REGULARIZATION 2
Individual-based Regularization
Treat each user’s friend differently based on the corresponding similarity
Social Network
A Large value of Sim(i,f) indicates a small distance between features
Both methods can be solved efficiently via gradient-based algorithm

11. SIMILARITY FUNCTION
Basic idea: make use of the observed ratings to compute the similarity between any two users
Pearson Correlation Coefficient (PCC)
centered version of cosine similarity

12. DATA AND STATISTICS Datasets Douban (Book/Movie/Music Review,
Rating, Social Network)
users, movies
ratings, friend links
Epinions (Consumer Review/Rating)
users, items
ratings, friend links

13. BASELINE METHODS UserMean ItemMean
RSTE: trust-aware recommendation algorithm.

14. EXPERIMENTAL RESULTS
Hide part of the observed ratings and estimate the hidden ones using different methods
Evaluate based on mean absolute error (MAE)
Method
UserMean
ItemMean
RSTE
SR1
SR2
Douban (80%)
0.6809
0.6288
0.5643
0.5576
0.5543
Epinion (90%)
0.9134
0.9768
0.8367
0.8287
0.8256

15. IMPACT OF SIMILARITY FUNCTION
Sim = Ran
Sim = PCC
Douban
0.5579
0.5592
0.5543
Epinion
0.8324
0.8345
0.8256

16. MAIN TAKEAWAY Conclusion:
Users' social friend information can improve recommendation accuracy
Future work:
Identify the most suitable group of friends to enhance recommendation accuracy
Weaknesses:
Social network in Douban may not be bilateral
Not using k-fold cross validation

17. LU TIAN, XIAO ZHANG, MENGMENG YE
Thank You!
RECOMMENDER SYSTEMS WITH SOCIAL REGULARIZATION HAO MA, DENGYONG ZHOU, CHAO LIU, MICHAEL R. LYU, IRWIN KING
LU TIAN, XIAO ZHANG, MENGMENG YE
MARCH 2018