Effective Heterophilous Heterogeneous Graph Learning via Decoupled Spectral Filtering and Curriculum Learning

Abstract

Heterogeneous graphs (HGs) are widely used to model complex relational data in domains such as traffic networks, biology, and recommendation systems. While heterogeneous graph neural networks (HGNNs) demonstrate strong performance on homophilous benchmarks, they struggle to generalize to heterophilous HGs where connected nodes often hold different labels. Recent spectral HGNNs attempt to address this by modeling complex shift operations by positive non-commutative polynomials, but they suffer from severe over-parameterization and overfitting, especially when data are limited. Furthermore, existing methods often neglect the internal conflict between meta-path subgraphs with diverging homophily ratios. In this paper, we propose CLS-HGNN, a Curriculum Learning enhanced Spectral Heterogeneous Graph Neural Network. To mitigate over-parameterization, we replace dense learnable weight matrices with Jacobi polynomial bases, which decouple the spectral filtering process and significantly reduce complexity while preserving expressiveness. To reduce the impact of conflicts arising from inconsistent homophily, we introduce a curriculum learning strategy that quantifies node difficulty via training loss dynamics and embedding deviations across meta-paths. This allows the model to progressively learn from consistent (easy) patterns to conflicting (hard) ones. Extensive experiments demonstrate that CLS-HGNN achieves state-of-the-art performance on heterophilous HGs and significantly improves training efficiency compared to existing spectral methods.