Learning Edge Representations via Low-Rank Asymmetric Projections

We propose a method for learning continuous-space vector representation of graphs, which preserves directed edge information. Previous work in learning structure-preserving graph embeddings learn one embedding vector per node. In addition to learning node embeddings, we also model a directed edge as a learnable function of node embeddings, which enable us to learn more concise representations that better preserve the graph structure. We perform both intrinsic and extrinsic evaluations of our method, presenting results on a variety of graphs from social networks, protein interactions, and e-commerce. Our results show that learning joint representations learned through our method significantly improves state-of-the-art on link prediction tasks, showing error reductions of up to 69% and 36%, respectively, on directed and undirected graphs, while using representations with 16 times less dimensions per node.