Abstract: The purpose of this study is to accurately identify the new emerging topic immediately and to identify its evolution law from all kinds of heterogeneous on-line social network data which are composed of short texts like news titles and micro-blogs, the results of which will provide valuable decision support for government officers and company administrators. Firstly, all kinds of heterogeneous on-line social network data are acquired and modeled as the time-varying network by using the co-occurrence of short texts' keywords, from which the topic early detection problem could be changed into the problem of dynamic community detection on the time-varying topic network; Secondly, we propose the dynamic community detection method with static Louvain algorithm embedded and with modularity gain and local network variation as quantification values. The proposed method  yields preferable results under both large amount of computer-generated data and real heterogeneous online social network data, community detection and propagation identification results under computer-generated and real heterogeneous online social network data validates the algorithms' efficiency and feasibility.

Key words: heterogeneous online social network, dynamic community detection, Louvain algo-rithm, modularity gain