MixHop: Higher-Order Graph Convolutional Architectures via Sparsified Neighborhood Mixing

Existing popular methods for semi-supervised learning with Graph Neural (such as the Graph Convolutional Network) provably cannot learn a class of neighborhood mixing relationships. To address this weakness, propose a new model, MixHop, that can learn these relationships, including operators, by repeatedly mixing feature representations of neighbors various distances. Mixhop requires no additional memory or computational, and outperforms on challenging baselines. In addition, we propose regularization that allows us to visualize how the network prioritizes information across different graph datasets. Our analysis of the architectures reveals that neighborhood mixing varies per datasets.