Over-smoothing problem of heterogeneous graph neural networks: A heterogeneous graph neural network with enhanced node differentiability

Abstract

Heterogeneous graph neural networks (HGNNs) are prone to node feature homogenization and over-smoothing issues during deep information propagation, leading to significant performance degradation. Through quantitative analysis using Dirichlet energy, this study reveals that traditional HGNNs exhibit a sharp decline in Dirichlet energy with increasing network depth, which impairs model performance. Furthermore, this paper provides the first rigorous theoretical proof, based on spectral analysis, of the feature degeneration mechanism in both graph convolutional aggregation and attention aggregation within deep architectures. To address these challenges, this paper propose a Heterogeneous Graph Neural Network with Enhanced Node Distinguishability (ENDHG). ENDHG introduces a novel graph convolution strategy to enhance the preservation of node-specific features during meta-path aggregation, thereby mitigating over-smoothing risks. Additionally, ENDHG effectively utilizes each node’s unique heterogeneous neighbor information to further improve node distinguishability. Compared to existing methods, ENDHG achieves approximately 2% performance improvement across three real-world datasets while demonstrating stronger over-smoothing resistance and more stable Dirichlet energy in deep architectures.