-
Why Spectral GNNs Excel by Discarding High-Frequency
A defining characteristic of spectral Graph Neural Networks (GNNs) is their inherent low-pass filtering mechanism, which involves discarding high-frequency signal components, a step that might initially appear 'arbitrary' or reductive. This blog explores why this seemingly reductive step of low-pass filtering, which prioritizes the smoother, low-frequency components, proves so beneficial for various node-level and graph-level tasks.