Abstract
Approximation-based spectral graph neural networks, which construct graph filters via function approximation, have demonstrated significant success in graph learning tasks. While existing methods predominantly rely on polynomial approximations for filter construction, rational approximation, a potentially superior alternative, remains largely underexplored. Previous attempts to utilize rational approximations have faced challenges, including high computational costs or reliance on polynomial intermediaries, limiting their full potential. This paper introduces TS-RatGNN, a novel spectral GNN featuring an explicitly-optimized rational filter. TS-RatGNN employs a distinctive two-stage framework: it sequentially applies numerator and denominator filters to the input signal. This streamlines the model architecture, facilitates efficient implementation, and uniquely enables the explicit optimization of both the numerator and the denominator components of the rational filter. Comprehensive experiments validate the effectiveness of TS-RatGNN compared to state-of-the- art methods, positioning it as a viable and practical approach for deploying rational filter-based GNNs.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
"TS-RatGNN: Two-Stage Spectral Graph Neural Network with Explicitly Optimized Rational Filters", Technical Disclosure Commons, (May 01, 2025)
https://www.tdcommons.org/dpubs_series/8065