Studies of social systems have traditionally focused on analyzing networks induced by social interactions, while discarding rich contextual information on nodes and their properties. At the same time, empirical evidence points to strong correlations between node attributes and their interactions.  Here we suggest a viable framework for analyzing attribute-rich and multi-modal social data based on latent space models. In this approach, each node is assigned an unobserved (latent) position in some space, so that both the nodes’ attributes and their interactions depend on their coordinates in this space. This “shared” latent space allows to capture observed correlations between the attributes and network structure. We perform extensive experiments  where the goal is predict missing links in a network using attributes, or predict user attributes based on network information, and observe that the proposed method outperforms other baselines in both prediction tasks.