Embedding networks with edge attributes

Abstract

Predicting links in information networks requires deep understanding and careful modeling of network structure. Network embedding, which aims to learn low-dimensional representations of nodes, has been used successfully for the task of link prediction in the past few decades. Existing methods utilize the observed edges in the network to model the interactions between nodes and learn representations which explain the behavior. In addition to the presence of edges, networks often have information which can be used to improve the embedding. For example, in author collaboration networks, the bag of words representing the abstract of co-authored paper can be used as edge attributes. In this paper, we propose a novel approach, which uses the edges and their associated labels to learn node embeddings. Our model jointly optimizes higher order node neighborhood, social roles and edge attributes reconstruction error using deep architecture which can model highly non-linear interactions. We demonstrate the efficacy of our model over existing state-of-the-art methods on two real world data sets. We observe that such attributes can improve the quality of embedding and yield better performance in link prediction.