A Comparison of Spectral and Spatial Graph Convolutional Neural Network Kernels Using GraphSAGE-Sparse

Graph Convolutional Networks (GCNs) are widely successful architectures for performing deep learning on graphs, but their well-known scalability challenges have led to increased interest to develop both improved algorithms and hardware accelerators. In this paper, we present and evaluate GraphSAGE-Sparse, a variant of the paradigmatic GraphSAGE GCN that replaces the original's spatial-based node convolution operation with a minibatch-aware sparse matrix multiply (SpMM) kernel. We find that this modification substantially reduces the per-batch memory cost for training and inference on a GPU accelerator, with the tradeoff of increased time and memory needed to preprocess the data structures used by the sparse kernel. On comparing both algorithms with datasets from the Open Graph Benchmark, we find that GraphSAGE-Sparse is able to obtain improved accuracy predictions in less than half of the total training time, even with the additional preprocessing work.