A spectral perspective on GCNs
In this work, we study the behavior of standard GCNs under spectral manipulations. We relate the numerical performances of a GCN through bandpass filtering: we empirically show that most of the necessary and used information for nodes classification is contained in the low-frequency domain, through various ablation experiments. In particular, it is possible to recover accuracies competitive with the state of the art with simple MLPs and only a few low frequencies. We further derive a simple procedure to adjust the trade-off between smoothing and high-frequency propagation for deeper GCNs, without diagonalizing the Laplacian. We conduct our experiments on standard and challenging benchmarks.
PDF Abstract