Abstract: This paper presents a novel micro-community detection method based on topic model. Multiple latent geographical regions by Dirichlet process are created adaptively. A multinomial distribution is then employed to depict topics evolutions within each time bin. User selection preferences of latent geographical region and community are introduced in topic model. Finally, the EM method and Gibbs sampling method are used to estimate spatio-temporal topic model parameters so that community detection can be realized by topics similarity. Experiment results show that this method can promote the performances of community identifying.