Parallel Message Passing in Dual-space on Graphs

Abstract

Graph neural networks (GNNs) have achieved a great success on graph-based tasks. Most existing GNNs pay attention to the message passing process, where features propagate along the topology of the graph. Recently, dimension-wise separable 2-D graph convolution (DSGC) points out that GNNs ignore the effect of interactions between the features, and specially designs a dimension-wise feature affinity matrix via predefined corpuses. Although DSGC further improves the performance, we find that it has two shortcomings. For one thing, it performs poorly on low homophily datasets. For another thing, it tends to be over-smoothing as the number of layers increases, making node embeddings indistinguishable. In our empirical analysis, these two shortcomings are caused by the serial message passing mechanism of topology and features. To solve this problem, we propose a parallel-space graph convolution (PSGC) in this paper. The central idea is that we design a parallel message passing mechanism in dual spaces including a topology space and a feature space, which can adapt to different homophily real-world datasets and alleviate the over-smoothing issue. Besides, we design the dimension-wise feature affinity matrix derived by the original node features instead of corpuses in DSGC to alleviate the severe overfitting in the feature space, where corpuses are not provided by many benchmark datasets. Our extensive experiments on benchmark datasets obviously show that PSGC achieves significant performance gain over state-of-the-art methods for semi-supervised node classification.