Seeing All From a Few: Nodes Selection Using Graph Pooling for Graph Clustering

Graph clustering aiming to obtain a partition of data using the graph information, has received considerable attention in recent years. However, noisy edges and nodes in the graph may make the clustering results worse. In this paper, we propose a novel dual graph embedding network(DGEN) to improve the robustness of the graph clustering to the noisy nodes and edges. DGEN is designed as a two-step graph encoder connected by a graph pooling layer, which learns the graph embedding of the selected nodes. Based on the assumption that a node and its nearest neighbors should belong to the same cluster, we devise the neighbor cluster pooling(NCPool) to select the most informative subset of vertices based on the clustering assignments of nodes and their nearest neighbor. This can effectively alleviate the impact of the noise edge to the clustering. After obtaining the clustering assignments of the selected nodes, a classifier is trained using these selected nodes and the final clustering assignments for all the nodes can be obtained by this classifier. Experiments on three benchmark graph datasets demonstrate the superiority compared with several state-of-the-art algorithms.