Existing popular methods for semi-supervised learning with graph neural networks (such as the Graph Convolutional Network) provably cannot learn a general class of neighborhood mixing relationships. To address th