Point-of-interest (POI) recommendation has become a favorite topic on location-based cyber-physical–social networks (LBCPSNs). The geographical influence of locations (e.g., geographical proximity) and the social influence of friends (e.g., social ties) have been widely used in many existing POI recommendation methods, most of which paid little attention to the sequential influence and temporal influence of locations on users’ check-in behaviors. However, human mobility exhibits sequential and temporal patterns. In this paper, we propose a unified probabilistic generative model, i.e., a multi-modal Bayesian embedding model (MMBE), to discover the social, sequential, temporal, and spatial patterns of users’ check-in behaviors simultaneously. Besides, MMBE can model the joint effect of the four factors mentioned above on the decision-making for POI selection across heterogeneous workflow activities in cyber-physical–social(CPS) systems. Then, we conduct an in-depth performance evaluation for MMBE using two large-scale real-world datasets from Gowalla and Brightkite. Experimental results show that our proposed method outperforms the other five state-of-the-art POI recommendation approaches regarding Precision and Recall. © 2017 Elsevier B.V.