Dual Feature Enhancement Graph Clustering Network

Abstract

Deep graph clustering is a fundamental method in unsupervised learning. Recently, deep clustering fusion methods relying on representation learning typically employ Auto-Encoders (AEs) and Graph Neural Networks (GNNs) to capture high-dimensional information representations of node attributes and graph structure. However, non-important graph structure information and redundant fused representations lead to a less discriminative graph representation, limiting clustering performance. To tackle this issue, we propose a Dual Feature Enhancement Graph Clustering Network (DFE-GCN). Specifically, we develop a critical node selection mechanism that calculates the importance score of each node to adjust edge weights, reducing non-important connections while enhancing important connections. Moreover, we design a heterogeneous information fusion strategy that fine-tunes the node attributes and graph structure fused layer by layer between nodes, dynamically filtering out redundant representations and forming a robust target distribution. Extensive experiments on five datasets have proven that the proposed method consistently outperforms advanced clustering methods.