Moving beyond graph homophily: Disentangled node representation for graph neural networks.

Abstract

Graph Neural Networks (GNNs) have demonstrated strong performance on graph-structured data through neighborhood aggregation. However, they often underperform on heterophilic graphs and suffer from over-smoothing, leading to suboptimal representations. We attribute these limitations to the lack of view-specific factorization in current node representations, which fail to reflect the multifaceted nature of node features. In this paper, we propose a novel node representation paradigm that decomposes node features into three complementary perspectives: label context, semantic, and topology. We observe that the relative importance of these perspectives varies not only across datasets but also among nodes within a single graph. To address this, we introduce RDGCN (Representation-Disentangled Graph Convolutional Network), which employs view-specific encoders and an adaptive fusion mechanism to learn and integrate multi-view representations dynamically. RDGCN mitigates harmful information mixing through view-specific factorization and enhances the discriminability of each encoder via targeted regularization. Notably, our framework is based on simple MLP and GNN modules yet achieves strong performance, demonstrating the power of multi-view representation learning. Moreover, our work establishes a generalizable node representation paradigm and offers a unified framework that moves beyond the binary homophily/heterophily divide. Extensive experiments on nine benchmark datasets validate the effectiveness of RDGCN, showcasing robust performance across both homophilic and heterophilic graphs.