Mineral: Multi-modal Network Representation Learning

Network representation learning (NRL) is a task of learning an embedding of nodes in a low-dimensional space. Recent advances in this area have achieved interesting results; however, as there is no solution that fits all kind of networks, NRL algorithms need to be specialized to preserve specific aspects of the networks, such as topology, information content, and community structure. One aspect that has been neglected so far is how a network reacts to the diffusion of information. This aspect is particularly relevant in the context of social networks. Studies have found out that diffusion reveals complex patterns in the network structure that are otherwise difficult to be discovered by other means. In this work, we describe a novel algorithm that combines topology, information content and diffusion process, and jointly learns a high quality embedding of nodes. We performed several experiments using multiple datasets and demonstrate that our algorithm performs significantly better in many network analysis tasks over existing studies.