RecSys’10: A Matrix Factorization Technique with Trust Propagation for Recommendation in Social Networks
Recommendation system paper challenge (4/50)
Why this paper?
RecSys’10 and top popular recommendation system paper.
What problem do they solve?
Recommendation system with social network and we can improve cold-user performance with social network information
They apply a model-based approach for recommendation in social networks with trust propagation.
Why social network would benefit?
The effects of selection and social influence can help the recommendation system.
Selection means that people tend to relate to people with similar attributes.
social influence related people in a social network influence each other to become more similar.
What do others solve the problem?
Memory based approaches: It is east to include trust/social effect into the approach. It use the transitivity of trust and propagate trust to indirect neighbors in the social network. Testing phase is slow.
Model based approaches: Usually model based approach is better memory based approach. However, no one include social effect into model-based yet.
STE model: It linearly combines the user-item information and neighboring rating behaviors. Neighboring rating directly impose user rating.
What is the solution?
SocialMF: They make the features of user dependent on the feature vectors of the his direct neighbors in the social network. It forces the user feature vectors to be close to those of their neighbors.
How do they measure the performance?
They want to predict the rating. Hence, they apply RMSE to measure the performance.
The data?
Epinons.com
Flixster.com
What is the baseline model?
BaseMF: This method is the model based matrix factorization approach, which does not take the social network into account.
CF: This is the user based collaborative filtering method which is a memory based approach.
STE: It takes into account the social network directly affecting to rating.
The difference between baseline models and STE?
STE does not support trust propagation and they list trust propagation as future work. STE just linearly combines the rating of neighbors. In SocialMF, its feature vector of each user is dependent on the feature vectors of his direct neighbors in the social network, which can handle the transitivity of trust and trust propagation.
baseMF and STE only model based on the observed rating data. However most of people did not rate yet participate in social activities. They loss the valuable information.
SocialMF
The behavior (latent vector) of a user u is affected by his direct neighbors but directly affect the user rating.
So the rating is still same.
For the user latent features, we have two factors: neighboring behavior (propagate social effect)+ Gaussian prior (avoid overfitting)
The result
Other related blogs:
Trust-aware recommender systems
Performance of recommender algorithms on top-n recommendation tasks
Best paper in RecSys:
https://recsys.acm.org/best-papers/
My Website: