Abstract
Spectral graph convolutional networks (SGCNs) have been attracting increasingattention in graph representation learning partly due to their interpretabilitythrough the prism of the established graph signal processing framework.However, existing SGCNs are limited in implementing graph convolutions withrigid transforms that could not adapt to signals residing on graphs and tasksat hand. In this paper, we propose a novel class of spectral graphconvolutional networks that implement graph convolutions with adaptive graphwavelets. Specifically, the adaptive graph wavelets are learned with neuralnetwork-parameterized lifting structures, where structure-aware attention-basedlifting operations are developed to jointly consider graph structures and nodefeatures. We propose to lift based on diffusion wavelets to alleviate thestructural information loss induced by partitioning non-bipartite graphs. Bydesign, the locality and sparsity of the resulting wavelet transform as well asthe scalability of the lifting structure for large and varying-size graphs areguaranteed. We further derive a soft-thresholding filtering operation bylearning sparse graph representations in terms of the learned wavelets, whichimproves the scalability and interpretablity, and yield a localized, efficientand scalable spectral graph convolution. To ensure that the learned graphrepresentations are invariant to node permutations, a layer is employed at theinput of the networks to reorder the nodes according to their local topologyinformation. We evaluate the proposed networks in both node-level andgraph-level representation learning tasks on benchmark citation andbioinformatics graph datasets. Extensive experiments demonstrate thesuperiority of the proposed networks over existing SGCNs in terms of accuracy,efficiency and scalability.