Multi-Graph Contrastive Learning Clustering Network

Abstract

Graph contrastive learning methods often construct numerous homogenous graphs by randomly adding edge perturbance and removing edges to/from one given graph, while it is difficult to guarantee the quality of these augmented graph structures. To fully guarantee and utilize the rich structural information provided by the above heterogeneous graphs, we propose a novel multi-graph contrastive clustering network model that collaborates various types of relationships between nodes. In contrast to traditional contrastive learning methods that only regard each node as one positive sample and its neighbors as negative samples within a single graph, we further build the contrastive constraint between nodes of the same sample in different graphs, improving the node representation capability from the topology and sample attribute aspects. The designed multi-graph attention mechanism assigns more weights to significant graphs and nodes. Experimental results on three public multi-graph datasets demonstrate that the proposed method achieves satisfactory performance compared to other state-of-the-art clustering methods.