Learning discrete adaptive receptive fields for graph convolutional networks

Summary: Graph Neural Networks (GNNs) are effective for representation learning on graph-structured data and have achieved state-of-the-art performance on various predictive tasks. Graph Attention Networks (GATs) improve the performance of graph learning tasks by assigning dense attention coefficients to node neighbors. However, GATs are prone to overfitting on large and noisy graphs, and may fail on disassortative graphs. Sparse Graph Attention Networks (SGATs) learn sparse attention coefficients using $L_0$-norm regularization, allowing feature aggregation on informative neighbors. SGATs outperform GATs on assortative and disassortative graphs and can remove noisy edges from assortative graphs while maintaining classification accuracies. This is the first graph learning algorithm that demonstrates redundancies in graphs and shows that edge-sparsified graphs can achieve similar or higher predictive performances than original graphs.