User profiling in an ego network

Abstract

User attributes, such as occupation, education, and location, are important for many applications. In this paper, we study the problem of profiling user attributes in social network. To capture the correlation between attributes and social connections, we present a new insight that social connections are discriminatively correlated with attributes via a hidden factor -- relationship type. For example, a user's colleagues are more likely to share the same employer with him than other friends. Based on the insight, we propose to co-profile users' attributes and relationship types of their connections. To achieve co-profiling, we develop an efficient algorithm based on an optimization framework. Our algorithm captures our insight effectively. It iteratively profiles attributes by propagation via certain types of connections, and profiles types of connections based on attributes and the network structure. We conduct extensive experiments to evaluate our algorithm. The results show that our algorithm profiles various attributes accurately, which improves the state-of-the-art methods by 12%.