With the rapid development of social networks in recent years, a large amount of short text data with time-spacial information is produced accordingly. Due to short length of text and sparseness of geographic location, it is very difficult to capture the semantic topics of user behavior. In addition, most existing research work related to user behavior understanding has not taken the behavior elements dependency into account, which results in the incomplete understanding of user behavior. Based on these, two models mixed with time, activity and region, i.e., user-time-activity model (UTAM) and user-time-region model (UTRM), are proposed firstly in this paper so as to explore behavior principles effectively. And then, by extracting activity-service topics based on latent Dirichlet allocation (LDA) techniques, an activity-to-service topic model (ASTM) is proposed in order to mine corresponding relationships between activities and services. Finally, a novel matrix factorization algorithm fused with distance and coupled similarity, i.e., matrix factorization based on couple & distance (MFCD), is put forward to improve the recommendation quality. In order to verify the effectiveness of proposed models and algorithms, extensive experiments are executed on a real Twitter dataset. Experimental results show that the proposed models can improve the quality of personalized recommendation service greatly, and the performance of MFCD algorithm is superior to the traditional matrix factorization algorithm on the effect of understanding user behaviors.