A social network-based recommender system

• Authors: Wesley W. Chu;Jianming He
• Venue: A social network-based recommender system
• Year: 2010

For more than a decade, recommender systems have been proposed as a key tool to overcome information overload, and many algorithms and systems have been developed. Despite all these efforts, recommender systems still face many challenges—such as improving prediction accuracy, data sparsity, and cold-start issues. On the other hand, the recent emergence of online social networks provides us with an unprecedented large amount of information regarding user behavior and friend interactions. Especially, homophily effects among friends have demonstrated their importance to product marketing, and marketing strategies that utilize homophily have achieved great success. However, such information has not yet been considered in recommender systems. In this dissertation we propose a new paradigm of recommender systems which can significantly improve performance by utilizing information in social networks. In particular, we address the following challenging issues in building a social network-based recommender system. 1) We investigate the existence of homophily among friends in rating items using statistical analyses on a dataset crawled from a real online social network, Yelp.com. 2) To understand the role of social relationships in a social network-based recommender system, we develop a Bayesian network-based recommender system (SNRS-BN) based on a simplified homogeneous social network, and study the performance of SNRS-BN under different types of social relationships. 3) We develop a social network-based recommender system (SNRS) which utilizes more information in social network, including user preferences, an item's likability, and homophily effects among friends, to provide better recommendations. An iterative classification has been adopted in this system to incorporate homophily even from friends at multiple hops away. The performance improvements of SNRS are validated through experiments on the Yelp dataset in terms of both prediction accuracy and coverage over traditional recommender systems. 4) To overcome the problems of heterogeneities in social networks, we propose to select relevant friends for inference based on the semantics in fine-grained user ratings on their buying decisions. An experiment is carried out in a graduate student class to validate performance improvements from such semantic filtering. 5) We discuss the trust issues caused by malicious users and users with unreliable domain knowledge in SNRS. In particular, for the case of users with unreliable domain knowledge, we propose to estimate homophily more robustly by relaxing the item category information according to item taxonomy. Future work in this direction will be desirable.