Multi-spectral attention and graph smoothness enhancement for generalized node classification

Area of Science:

• Graph Neural Networks
• Spectral Graph Theory
• Machine Learning

Background:

• Existing spectral Graph Neural Networks (GNNs) face limitations due to fixed polynomial bases and predefined propagation mechanisms.
• These limitations hinder adaptability to real-world graphs with complex spectral properties and varying homophily levels.

Purpose of the Study:

• To propose UniSpecAR, a Unified Spectral Adaptive Representation framework, to overcome the limitations of current spectral GNNs.
• To introduce adaptive mechanisms for dynamic construction, fusion, and balancing of spectral and spatial information.

Main Methods:

• Developed a novel Krylov multi-basis filter for dynamic construction and fusion of information across diverse frequency subspaces.
• Implemented parallel diffusion channels with a channel-level attention mechanism for adaptive fusion.
• Introduced a spatial-spectral gating mechanism to balance spectral features with local topological structures.
• Incorporated a spectral consistency regularizer to ensure filter stability and structural faithfulness by penalizing complexity and deviation from local topology.

Main Results:

• UniSpecAR significantly outperforms state-of-the-art GNNs on multiple benchmarks, demonstrating effectiveness on both homophilic and heterophilic graphs.
• The adaptive multi-basis design and fusion mechanisms enhance spectral expressiveness.
• The framework provides valuable interpretability of learned spectral features.

Conclusions:

• UniSpecAR offers a unified and adaptive approach to spectral representation in GNNs.
• The proposed methods effectively address the adaptability challenges posed by complex graph structures and varying homophily.
• The framework demonstrates superior performance and interpretability, advancing the field of spectral GNNs.