Graph-Enforced Neural Network for Attributed Graph Clustering

Graph clustering aims to discover cluster structures in graphs. This task becomes more challenging when each node in the graph is associated with an attribute vector (i.e., the attributed graph). Recently, methods built on Graph Auto-Encoder (GAE) have achieved state-of-the-art performance on the attributed graph clustering task. The performance gain mainly comes from GAE's ability to capture knowledge from graph structures and node attributes simultaneously. However, there is limited understanding of the critical issues that hinder the clustering performance of current GAE-based methods. To bridge this gap, we present a detailed empirical analysis and find that existing GAE-based methods suffer from graph information degradation issues of intra-cluster estrangement, attribute similarity neglection, and blurred cluster boundaries. Based on the observations, we design corresponding graph-enforcement tasks to address these degradation issues and include them in a unified multi-task learning framework called Graph-Enforced Neural Network (GENN). Extensive experimental results on four popular graph benchmark datasets illustrate that GENN consistently outperforms state-of-the-art attributed graph clustering methods.