Recommender System with Composite Social Trust Networks

Introduction

In an era of social connectedness, people are becoming increasingly enthusiastic about interacting, sharing, and collaborating through social networks, online communities, blogs, wikis, and other online collaborative medias (Cambria, Rajagopal, Olsher, & Das, 2013). Big social data has driven researchers to pay more and more attention to the improvement of recommender algorithms in order to provide an intelligent mechanism for filtering the excess information available to users (Park & Chang, 2009; Choi, Yoo, Kim, & Suh, 2012). As an efficiency way to information filtering, recommender systems have been studied and deployed extensively over the last decade in various application areas, including e-commerce, social networks, and advertisements.

Collaborative filtering (CF) (Breese, Heckerman, & Kadie, 1998) is the most popular recommender system, which is based on the simple assumption that similar users tend to share similar interests. CF has attracted considerable attention since its effectiveness, and has been developed and adopted by large, successful commercial systems, including Amazon¹ (Linden, Smith, & York, 2003) and Netflix².

However, traditional CF technology suffers from several inherent weaknesses, such as data sparsity, imbalance of rating data, and shilling attacks (Su & Khoshgoftaar, 2009). Recently, online social networks such as Epinions, Twitter, and Lastfm, where users can link with each other and share content (such as comments, messages, and music), have become popular platforms. To overcome the above mentioned problems, trust-based recommender systems have been studied in social recommendations. Trust-based recommender systems incorporate social trust information to do recommendation, and thus can mitigate the data sparsity problem by capturing social trust relationships that are stored outside each user’s local similarity neighborhood (Lathia, Hailes, & Capra, 2009). Social trust information also makes recommender systems resistant to shilling attacks to some extent, by preventing malicious community members from abusing the system. Previous studies (Bedi & Sharma, 2012; Chen, Zeng, Zheng, & Chen, 2013; Yan, Zheng, Chen, & Wang, 2013) have shown that trust-based recommendations always outperform traditional CF algorithms in terms of prediction accuracy and decision-support accuracy.

Researchers generally adopt two approaches when establishing trust networks: explicit methods that draw explicit social trust networks from pre-established (or manually input) social links among users, and implicit methods that infer implicit social trust among users on the basis of their buying, rating, or other interaction histories (Lathia, Hailes, & Capra, 2009). Both methods assume that underlying relationships (either pre-existing or inferred) can be described and rationalized as a web of trust. Webs of trust are conceptualized as graphs in which users are nodes and edges are weighted according to the extent that users trust each other (Lathia, Hailes, & Capra, 2009).

However, most existing research focuses either on explicit social trust (Jamali & Ester, 2010) or simply using implicit social trust (Ma, Zhou, Liu, Lyu, & King, 2011; Liu & Aberer, 2013). None of the approaches builds a rational social trust network based on both users’ explicit trust and implicit trust, thus causes information lose and defects recommendation performance. Moreover, most existing work either minimizes the difference between the true ratings and the inferred ratings (Ma, King, & Lyu, 2009), or minimizes the difference between the real trust values and the inferred trust values (Ma, Yang, Lyu, & King, 2008). Chen et al. (2013) minimize both the rating difference and the trust difference between the true values and the inferred values, but use only users’ explicit social trust networks, ignoring users’ implicit social trust networks.