Publication Type

Journal Article

Version

publishedVersion

Publication Date

3-2026

Abstract

Graph Neural Networks (GNNs) face two key challenges, heterogeneity and heterophily, which often degrade performance. Existing approaches either focus narrowly on specific meta-paths, limiting their expressiveness, or are expressive but cannot effectively leverage higher-order neighbors. In this paper, we propose the Heterogeneous Heterophilic Spectral Graph Neural Network (H2SGNN), which combines local independent filtering to adaptively handle meta-path subgraphs with varying homophily ratios, and global hybrid filtering to capture high-order neighbor interactions with linear computational complexity. On five heterogeneous graph benchmarks—DBLP, ACM, IMDB, AMiner, and Yelp—H2SGNN consistently outperforms strong baselines, for example, achieving +1.0% Macro-F1 and +1.3% Micro-F1 on IMDB. It uses 3–6× fewer parameters, requires about one-quarter of the GPU memory, and reduces training time by up to 8× on AMiner. These results demonstrate that H2SGNN effectively addresses both heterogeneity and heterophily, achieving higher accuracy while significantly reducing parameter count and memory usage.

Keywords

Spectral graph neural networks, Heterophilic graphs, Heterogeneous graphs

Discipline

Artificial Intelligence and Robotics | Databases and Information Systems

Research Areas

Intelligent Systems and Optimization

Areas of Excellence

Digital transformation

Publication

Neurocomputing

Volume

683

First Page

1

Last Page

15

ISSN

0925-2312

Identifier

10.1016/j.neucom.2026.133500

Publisher

Elsevier

Additional URL

https://doi.org/10.1016/j.neucom.2026.133500

Download

Share

COinS