Reproduce “HeroFilter: Adaptive Spectral Graph Filter for Varying Heterophilic Relations” on OpenResearch

About this paper

Shuaicheng Zhang, Haohui Wang, Junhong Lin, Xiaojie Guo, Yada Zhu, Si Zhang, Dongqi Fu, Dawei Zhou

HEROFILTER, a collaboration from Virginia Tech, MIT, IBM Research, and Meta AI, introduces an adaptive spectral graph filter and MLP-Mixer architecture designed to address the challenges of varying heterophily in graph neural networks. The method consistently achieves superior node classification accuracy across diverse datasets, showing up to 9.2% accuracy improvement on heterophilic graphs while maintaining strong performance on homophilic ones.

Key insights

• The relationship between graph heterophily and optimal spectral filter responses is complex and non-monotonic across frequency components, necessitating adaptive filtering rather than fixed low-pass or high-pass approaches.
• Adaptive polynomial filters possess sufficient expressive power to align with arbitrary label distributions in the frequency domain, enabling the learning of optimal filter responses.
• Constructing spectrally relevant neighborhoods (patches) for each node, rather than relying solely on topological adjacency, is crucial for effectively aggregating information in heterophilic graphs.

Summary by alphaXiv.