AEGCN: An Autoencoder-Constrained Graph Convolutional Network

We propose a novel neural network architecture, called autoencoder-constrained graph convolutional network, to solve node classification task on domains. As suggested by its name, the core of this model is operating directly graphs, whose hidden layers are constrained an autoencoder. Comparing with vanilla networks, autoencoder step added reduce information loss brought Laplacian smoothing. consider applying our both homogeneous graphs and heterogeneous graphs. For approximates adjacency matrix input taking layer representations as encoder another one-layer decoder. since there multiple matrices corresponding different types edges, feature instead, changes particularly designed multi-channel pre-processing two layers. In cases, error occurred in approximation goes penalty term function. extensive experiments citation networks other we demonstrate that adding constraints significantly improves performance networks. Further, notice technique can be applied attention improve well. This reveals wide applicability proposed technique.