Graph Convolutional Networks Using Node Addition and Edge Reweighting

Graph convolutional networks (GCNs) provide a promising way to explore datasets that have graph structures in nature. The presence of corrupted or incomplete graphs, however, dramatically decreases the performance of GCNs. To improve the performance, recent works on GCNs reweighted edges or added missing edges on the given graphs. On top of that, this paper further explores the domain of node addition. This paper presents a simple but effective extension of GCNs by combining node addition and edge reweighting. Node addition adds new nodes and edges as communication centers to the original graphs. By doing so, nodes can share information together for efficient inference and noise reduction. Moreover, edge reweighting re-distributes the weights of edges, and even removes noisy edges considering local structures of graphs for performance improvement. Based on four publicly available datasets, the experimental results demonstrate that the proposed approach can achieve better performance than four state-of-the-art approaches.