Trust Inference for Social Networks

The online social networks such as Facebook and Twitter are becoming important platforms for making social connections and sharing information. Meanwhile, the powerful personal mobile computing devices, e.g., smart phones and tablets, and the ubiquitous wireless communications enable users to participate in on-line social activities and process complicated social media data including texts, videos, images, and voices anywhere any time. Hence, social computing applications have attracted significant interest from both industry and government. However, due to the open nature of such social networking platforms, they are inherently vulnerable to malicious spam users who misuse social networks to perform spamming, phishing, and spreading false information. Such misbehaviors can greatly compromise the utility and functionality of social networks. Therefore, there is an urgent need to develop effective measures to evaluate the trust of users in social computing networks.

Most existing approaches focus on discovering patterns and features of users, and use them as input to classification machine learning tools, e.g., support vector machines, for automatic trust analysis. But such classification approaches determine the trust of each user separately from other users. However, users in social networks are naturally related to each other through social connections and interactions, and users share common ground usually exhibit close relationships. For example, recently the case study result on the Twitter social network has revealed that criminal user accounts tend to be socially connected to form a small-world network.

We develop an effective and efficient probabilistic inference framework for solving the trust inference problem in social computing networks. In particular, we formulate the problem in a Pairwise Markov Random Field (PMRF) that takes into count both user features and user relationships. PMRF is a probabilistic graphical model that expresses the factorization of a global function, and hence we can apply the Belief Propagation (BP) algorithm to perform inference efficiently. BP exploits the graph structure for computing marginal functions from global functions of many variables.