PageRank Centrality based Graph Convolutional Networks for Semi-supervised Node Classification

Graph convolutional neural networks(GCNs) have achieved impressive results on a variety of graph-based tasks. However, due to the problem of over-smoothing, the output representation of each node in the graph converges to the same asymptotically with increasing depth of GCN, so the vast majority of GCN models in practical applications are shallow. In this paper, we investigate deep GCNs and propose a PageRank-GCN(PR-GCN) model, which is an extension of DropEdge model. PR-GCN uses PageRank to get the top e neighbors that are most easily accessed by nodes and adds them back to the adjacency matrix after DropEdge processing. The results of a series of experiments on real-world datasets show that the PR-GCN model outperforms the state-of-the-art baseline approach in semi-supervised node classification tasks. The experiments also validate that the proposed method can prevent over-smoothing in deep GCNs while mitigating the information loss caused by DropEdge.