Topic-level Community Detection

With the rapid development of on-line social networks (OSNs), people are spending more and more time in social networks. It is no doubt that OSNs have become an indispensable part in people’s daily lives. However, how to provide better services for OSN users and explore useful commercial information from OSN data, has become a critical problem that OSN providers need to address. Driven by this necessity, the Community Detection problem has received significant attention from both academia and industry. In OSNs, a community (also called clique) is defined as a group of users in which the users mutually have strong social correlations. Community detection is of great importance as we can partition users into different groups and recommend for each group activities of interest.Existing community detection methods are mainly based on social ties in OSNs. They employ graph theory techniques to find dense subgraphs, and rely on the user-created tags to discover the activities of interest for each group. However, such a method may suffer from the following drawbacks:(1)Low efficiency:each community structure is typically a small subgraph in the entire social network, searching the huge social network for each community could be rather inefficient, render the method difficult to work in distributed systems;(2)Data sparsity:as the user-tag matrix is quite sparse and two different tags may be synonym, analyzing user interest for each group directly based on tags may be inaccurate;(3)Structural overlapping:the partition process is merely based on social relationships between users and thus ignores the fact that each individual may belong to multiple groups.To address the above problems, we propose a novel community detection schema named topic-level community detection in this paper. Our method employ topic modeling methods to construct a topic profile for each user, and adapts existing community detection techniques to efficiently find community structures for each topic. Extensive experiments confirm that our solution could effectively solve the problems of existing methods.