Stochastic Training of Graph Convolutional Networks with Variance Reduction

Control variate based algorithms which allow sampling an arbitrarily small neighbor size are developed and a new theoretical guarantee for these algorithms to converge to a local optimum of GCN is proved.