Tensor Decomposition for Spectral Graph Convolution: Improving Expressivity and Performance

The authors propose a general form of spectral graph convolution that represents the filter coefficients as a third-order tensor. They then derive two novel spectral graph convolution architectures, CoDeSGC-CP and CoDeSGC-Tucker, by performing CP and Tucker decomposition on the coefficient tensor, respectively. These models achieve favorable performance improvements over state-of-the-art methods on various real-world datasets.