Research Of Topic Tracking Algorithm In Dynamic Social Network

Dynamic social network stream can be described as a result of the interactionbetween time-varying network structure and text documents. A typical example isacademic collaboration network, in which the collaboration between authors and paperspublished by authors are always changing over time. Another example is Microblogsystem, in which what the users generate and the relationship between users also changesdynamically.One of important tasks in dynamic social network analysis is how to find and trackuser-interested topic over time. At present, there are many works studying network streamor text document stream separately, however, there are less research which take intoaccount both factors and their interaction. In this paper, in order to track topics over time,we formally define the problem and propose a novel statistical method that models theinterplay between textual content and network structures.In order to track the latent topics, we propose a novel statistical model consideringthe social influence, selection phenomenon, historical status and textual content. Ourmodel is called Dynamic-social-network based Topic-Tracking model, short for DTT.Specifically, user’s interest in topics at current time is influenced by his own and hisneighbors’ interests, a Gibbs Random Field is defined to model this influence. The linkrelationships between users are influenced by each user’s interest in various topics. Weadopt a method that is similar to ’community impact link’ to model the selectionphenomenon. Thereafter a classic topic model is used to describe the generative process ofthe documents. Finally, in order to ensure the continuity of the topic itself, the dependenceof the historical status is also taken into account. Based on the above considerations, theformal definition of the topic tracking problem in dynamic social networks is given, andan effective algorithm for model learning is proposed.Based on the proposed model, experiments are performed on an academiccollaboration network data set of the Arnetminer. The experimental results show that thetopic tracking model has a good topic tracking performance, better model generalizationcapability, and outperforms existing methods in various perspectives. Furthermore, we qualitatively and quantitatively demonstrate that topics detected by Dynamic-social-network based Topic-Tracking model are more interpretable and semantically coherentthan those detected by traditional topic tracking models.