Abstract
This paper aims at revisiting Graph Convolutional Neural Networks by bridgingthe gap between spectral and spatial design of graph convolutions. Wetheoretically demonstrate some equivalence of the graph convolution processregardless it is designed in the spatial or the spectral domain. The obtainedgeneral framework allows to lead a spectral analysis of the most popularConvGNNs, explaining their performance and showing their limits. Moreover, theproposed framework is used to design new convolutions in spectral domain with acustom frequency profile while applying them in the spatial domain. We alsopropose a generalization of the depthwise separable convolution framework forgraph convolutional networks, what allows to decrease the total number oftrainable parameters by keeping the capacity of the model. To the best of ourknowledge, such a framework has never been used in the GNNs literature. Ourproposals are evaluated on both transductive and inductive graph learningproblems. Obtained results show the relevance of the proposed method andprovide one of the first experimental evidence of transferability of spectralfilter coefficients from one graph to another. Our source codes are publiclyavailable at: https://github.com/balcilar/Spectral-Designed-Graph-Convolutions