A Framework for Accelerating Graph Convolutional Networks on Massive Datasets

Xiang Li, Ruoming Jin, Rajiv Ramnath, Gagan Agrawal

Abstract: In recent years, there has been much interest in Graph Convolutional Networks (GCNs). There are several challenges associated with training GCNs. Particularly among them, because of massive scale of graphs, there is not only a large computation time, but also the need for partitioning and loading data multiple times. This paper presents a different framework in which existing GCN methods can be accelerated for execution on large graphs. Building on top of ideas from meta-learning we present an optimization strategy. This strategy is applied to three existing frameworks, resulting in new methods that we refer to as GraphSage++, ClusterGCN++, and GraphSaint++. Using graphs with order of 100 million edges, we demonstrate that we reduce the overall training time by up to 30%, while not having a noticeable reduction in F1 scores in most cases.