SOCIAL RECOMMENDATION SYSTEMS WITH ITEM-BASED REGULARIZATION

MPhil Thesis Defence


Title: "SOCIAL RECOMMENDATION SYSTEMS WITH ITEM-BASED REGULARIZATION"

By

Mr. Hao GE


Abstract

Given the increasingly large amount of information on the Internet, 
recommendation systems have been widely used to recommend interesting 
information to the users so that they do not have to proactively look for 
it by using search engines or visiting websites. Recommendation systems 
can help users filter out irrelevant information, thus reducing the 
information overloading problem. Further, by monitoring users actions on 
the recommendations, information producers can gain a better understanding 
of the users’ interest and hence can focus their resources on delivering 
information that would arouse the users interest.

In parallel with the advent of recommendation systems, online social media 
has gained dramatic increase in usage across different communities. The 
social relations created in social networks are of great value in 
improving the performance of traditional recommendation systems because 
people who are socially related have strong influence on each other in 
their interests, tastes and opinions, etc. The synergy between 
recommendation systems and social networks has received much attention 
from researchers in both industry and academia.

In this thesis, we develop a new technique to integrate social information 
into recommendation systems to improve recommendation quality. We choose 
matrix factorization model as the basic recommendation framework upon 
which we add social information. In matrix factorization model, both users 
and items are mapped into a joint latent factor space and the user-item 
ratings are inner products of vectors in that space. To take advantage of 
social information, we introduce regularization terms to influence the 
result of matrix factorization. Previous methods typically use user-based 
regularization term to make recommendations of socially connect users as 
close to each other as possible. Unfortunately, it is hard to identify 
users with similar interests because a user may have multiple interests 
and having some identical interests does not mean that the users have 
identical interests overall. Instead of using user-based regularization 
term, our proposed method introduces item-based regularization term. The 
advantage of our approach is that comparing to the diversity of user 
interest, an item typically has specific purposes, properties or 
applications, making it easier to identify similar items from the users 
who are interested in them. For example, a user is interested in a book on 
Data Mining and another user is interested in a book on Machine Learning, 
we can infer that they have some similar interests but cannot conclude 
that they have identical interests because they may have different 
interests in other domain (e.g., movie). On the other hand, there should 
be much overlap between users who are interested in a book on Data Mining 
and users who are interested in a book on Machine Leaning. Thus, the two 
books should be similar and share close latent vectors. We conduct 
experiments using a dataset from Douban, and Mean Average Error and Root 
Mean Square Error are used as performance metrics. We demonstrate that our 
method can improve the performance in the recommendation tasks.


Date:			Friday, 15 April 2016

Time:			10:30am - 12:30pm

Venue:			Room 2130B
 			Lift 19

Committee Members:	Prof. Dik-Lun Lee (Supervisor)
 			Dr. Raymond Wong (Chairperson)
 			Dr. Wilfred Ng


**** ALL are Welcome ****