Graph Representation Learning via Diversity-preserving Graph Refinement

For real-world graph data, the complex relationship between nodes is often represented as a hard binary link. Obviously, it is a discrete and simplified form of continuous relationship between nodes, which seriously limits the expressibility of the learned node representation. On the other hand, the node representation obtained in the embedding space can in turn be used to reveal the intrinsic relationship between nodes. To better characterize the node relationships and further facilitate the learning of node representation, an intuitive way is to refine the originally given graph structure with the embedded node representations. However, such global refinement of the relationships among all nodes without distinction will inevitably lead to some noisy edges, which may further confuse the training of the node representation learning model. In addition, it also has scalability problems on large graphs. To address these issues, we propose a local structure aware graph refinement to progressively reveal the latent relationships of nodes, thus achieving efficient and robust graph refinement.