Mining Social Information for Recommender System

Abstract

With the development of e-commerce and mobile commerce, the recommender system has become more and more important both in research and in practice. Traditionally, those people researching recommender system make use of user rating lists as the basis for making recommendations, while at the same time ignoring other factors which could affect users’ preferences. However, the fact is that user preferences and purchase decisions depend on both the purchaser’s personal experience and the prevailing social information available. Previous recommendation studies have taken social information into consideration when they estimate both user preferences and user ratings. However, these studies only focus on user behavior which may create privacy problems, or they simply take note of the distance between users to measure the strength of social influence on decision makers. This research focuses on how to make full use of the social information available in order to make more accurate recommendations. The term “social information” refers to information from the social environment, which not only includes information and recommendations from the close friends of users, but also includes information from widely available forums which contain public comments. According to the theoretical foundations learned from previous social, behavioral and recommendation studies, we propose three novel social recommendation methods. The field experiments described in this research compare the accuracy between existing recommendation methods (including the classical, traditional recommendation methods; recent classical recommendation methods, and previous social recommendation methods), and social recommendation methods which takes into account information mined from friends, as well as social recommendation methods which relies on information mined from public opinion forums, and finally the social recommendation methods which take into account information gathered from both friends and public opinion forums. The results of our experiment indicate that our three proposed social recommendation methods could achieve much higher levels of performance in terms of recall than previous recommendation methods. This is especially true when users do not provide any ratings to movies. In those cases, social recommendation methods become even more irreplaceable, because the other tried and tested recommendation methods simply cannot work under these conditions. Our research also indicates that the social information mined from friends plays a more important role in the estimation of user preferences and ratings than does the social information obtained from public forums.

Date of Award	4 Nov 2013
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Shaoyi Stephen LIAO (Supervisor)