Point-of-Interest recommendation is an efficient way to explore interesting unknown locations in social media mining of social networks. In order to solve the problem of sparse data and inaccuracy of single user model, we propose a User-City-Sequence Probabilistic Generation Model (UCSPGM) integrating a collective individual self-adaptive Markov model and the topic model. The collective individual self-adaptive Markov model consists of three parts such as the collective Markov model, the individual self-adaptive Markov model and the self-adaptive rank method. The former determines the topic sequence for all users in system and mines the behavioral patterns of users in a large environment. The later mines behavioral patterns for each user in a small environment. The last determines a self-adaptive-rank for each user in niche. We conduct a large amount of experiments to verify the effectiveness and efficiency of our method.