AdjMix: simplifying and attending graph convolutional networks

Abstract Simple graph convolution (SGC) achieves competitive classification accuracy to convolutional networks (GCNs) in various tasks while being computationally more efficient and fitting fewer parameters. However, the width of SGC is narrow due over-smoothing with higher power, which limits learning ability representations. Here, we propose AdjMix, a simple attentional model, that scalable wider structure captures nodes features information, by simultaneously mixing adjacency matrices different powers. We point out key factor mismatched weights matrices, design AdjMix address GCNs adjusting matching values. Experiments on citation including Pubmed, Citeseer, Cora show our improves over 2.4%, 2.2%, 3.2%, respectively, achieving same performance terms parameters complexity, obtains better accuracy, parameters, compared other baselines.