On the Experimental Transferability of Spectral Graph Convolutional Networks

Spectral Graph Convolutional Networks (GCNs) are generalisations of standard convolutional for graph-structured data using the Laplacian operator. Recent work [26] has shown that spectral GCNs have an intrinsic transferability. This work verifies this by studying the experimental transferability of spectral GCNs for a particular family of spectral graph networks using Chebyshev polynomials. This work introduces two contributions. First, numerical experiments exhibit good performances on two graph benchmarks [13] [19], on tasks involving batches of graphs, namely graph regression, graph classification and node classification problems. Secondly we study a form of data augmentation through structural edge dropout showing performance improvements for GCNs. This work contributes to open research with public implementations of all experiments, enabling full reproducibility.