FreLinear: spectral-aware design and acceleration for efficient graph neural networks

Abstract

Graph Neural Networks (GNNs) excel in modeling graph-structured data but often face significant computational costs and fail to capture high-frequency components critical for fine-grained local variations. We propose FreLinear, a novel framework that integrates spectral-domain analysis with an efficient linear-attention mechanism. By avoiding the quadratic complexity inherent in traditional Transformer architectures, FreLinear leverages Fourier-based spectral features to enhance sensitivity to local structures while achieving near-linear computational complexity. Extensive experiments across diverse benchmark datasets demonstrate that FreLinear consistently surpasses state-of-the-art GNNs, delivering superior accuracy with significantly reduced computational overhead. On eight public datasets such as arxiv and Citeseer, the running time was shortened by 1 to 3 times with an increase in the number of parameters. At the same time, on the node classification task, the performance was improved by an average of 1.4 percentage points compared to the previous best work in these eight datasets. The code for the method proposed in our paper is publicly available on https://github.com/SWLee777/Frelinear.