Title:  Social Computing for Personalization and Credible Information Mining using Probabilistic Graphical Models

Committee: 

Dr. Fekri, Advisor        

Dr. Davenport, Chair

Dr. McLaughlin

Abstract: The objective of the proposed research is to study social computing for personalized recommender systems and credible information mining. Collaborative Filtering (CF) is the most popular recommendation algorithm, which exploits historic user data, e.g., past ratings, to predict unknown user preferences. We develop probabilistic graphical models and inference algorithms for CF recommender systems. We propose to compute user similarity on appropriately chosen factor graphs, and model unknown ratings via Pairwise Markov Random Fields (PMRFs). We apply the Belief Propagation (BP) algorithm which exploits the factorization encoded in graphical structures to perform inference efficiently.  Moreover, CF algorithms are faced with security and privacy challenges. To counter shilling attacks, we develop a probabilistic factor graph model that exploits the collaborative spamming behaviors among spammers to jointly detect them. To preserve user privacy, we develop a semi-distributed item-based CF system using BP, without directly disclosing user ratings to the server or other peer users. Meanwhile, online social media services like Twitter are widely adopted by people all around the world. However, social media is also increasingly exploited to spread rumors and false information as well as to carry out malevolent activities such as spamming and phishing. We develop a generative probabilistic model for event credibility prediction in social media, and propose an online prediction algorithm based on streaming messages like tweets in Twitter. To predict social media user trustworthiness, we propose a probabilistic PMRF model that takes into account both user features and social relationships, where users with close social relationships are more likely to have similar trustworthiness.