In online social media, social action prediction and social tie discovery are two fundamental tasks for social network analysis. Traditionally, they were considered as separate tasks and solved independently. In this paper, we investigate the high correlation and mutual influence between social actions (ie user-behavior interactions) and social ties (ie user-user connections). We propose a unified coherent framework, namely mutual latent random graphs (MLRGs), to flexibly encode evidences from both social actions and social ties. We introduce latent, or hidden factors and coupled models with users, users’ behaviors and users’ relations to exploit mutual influence and mutual benefits between social actions and social ties. We propose a gradient based optimization algorithm to efficiently learn the model parameters. Experimental results show the validity and competitiveness of our model, compared to several state-of-the-art alternative models.