FastGCN: Fast Learning with Graph Convolutional Networks via Importance Sampling

The graph convolutional networks (GCN) recently proposed by Kipf and Welling an effective graph model for semi-supervised learning. This model, however, originally designed to be learned with the presence of both training and data. Moreover, the recursive neighborhood expansion across layers poses and memory challenges for training with large, dense graphs. To relax the of simultaneous availability of test data, we interpret graph as integral transforms of embedding functions under probability. Such an interpretation allows for the use of Monte Carlo approaches consistently estimate the integrals, which in turn leads to a batched scheme as we propose in this work---FastGCN. Enhanced with importance, FastGCN not only is efficient for training but also generalizes well inference. We show a comprehensive set of experiments to demonstrate its compared with GCN and related models. In particular, training is of magnitude more efficient while predictions remain comparably.