Learning to Recommend with Explicit and Implicit Social Relations
Introduction:
For this blog assignment, I summarized an interesting academic paper I found using Google Scholar. I start by providing basic information about the paper, then state why I found it interesting, and proceed to note the basic idea discussed, the assumptions considered, and finally provide an overview.
Title:
Learning to recommend with explicit and implicit social relations
Authors:
Hao Ma (The Chinese University of Hong Kong)
Irwin King (The Chinese University of Hong Kong)
Michael R. Lyu (The Chinese University of Hong Kong)
Article:
http://dl.acm.org/citation.cfm?id=1961201
Interesting:
The research suggests that a user’s trusted friends on the web affect that user’s online behavior. This is not a new idea, as most people would intuitively agree with this statement, but the authors show strong supporting evidence of this claim through a data-centric framework.
Basic Idea:
The authors term “Social Trust Ensemble,” or RSTE, as the fusion of the User-Item Matrix & Trust-Aware Systems. The User-Item matrix is considered explicit relations, as these refer to concrete instances where a user directly liked an object, or selected a rating for an object on some some predefined scale. Trust is considered an implicit relation, as the user did not directly provide any input that could be used to determine whether or not they like an object; this information is inferred from a user’s “trusted” friends. The problem studied in this article is how to predict the missing values for users effectively and efficiently by employing the trust graph and the user-item rating matrix.
Assumptions:
1.) Users have their own characteristics, and they have different tastes on different items, such as movies, books, music, articles, food, etc.
2.) Users can be easily influenced by the friends they trust, and prefer their friends’ recommendations.
3.) One user’s final decision is the balance between his/her own taste and his/her trusted friends’ favors.
*Strictly from my own experiences, I feel that these are valid assumptions.
Overview:
This academic paper focuses on a systematic approach towards understanding the relationship between the trust network and the user-item matrix. Two recommender systems are created; one utilizes only the user-item matrix to make recommendations, while the other uses only a trust-based system. After the performances of these two systems are individually analyzed, a parameter alpha is used to fuse these two systems into one, which the authors’ term: Social Trust Ensemble. If alpha = 1, the system only mines the User-Item matrix for matrix factorization; if alpha = 0, the system only extracts information from the social trust graph. The system’s optimal performance occurs at alpha = .6 (i.e., more weight allocated in favor of the Social Trust Graph, as opposed to the User-Item Matrix ). This is interesting because it suggests that a user’s trusted friends on the Web affect this user’s online behavior.
The Social Trust Ensemble shows to be useful for overcoming data scarcity (the authors’ algorithm consistently performs better than other methods, especially when few user ratings are given) and scalability (scales linearly with the number of observations, so it can be applied to large data sets) problems. The performance of the authors’ algorithm/system is compared with seven other approaches: UserMean (mean value of every user), ItemMean (mean value of every item), NMF (only the user-item matrix), PMF (probabilistic matrix factorization), TCF (trust-aware collaborative filtering), Trust (only using trusted friends’ tastes), and SoRec(social trust-aware recommendation method that factorizes the user-item rating matrix and users’ social trust network by sharing the same latent space) using the metrics MAE (Mean Absolute Error) and RMSE (Root Mean Square Error). The authors do not consider trust propagation and distrust; these are two possible extensions of the framework these authors have developed. These are two very real aspects observed in everyday life when providing a recommendation, and as such, it would be very interesting to see the impact that these two scenarios would have on the performance of this recommendation system.