TRC-GCN: Two Residual Connections in Graph Convolution Networks for Node Classification Tasks

Recently, several variants of graph convolution networks (GCNs), which have shown awesome performance in node classification tasks, were investigated. This paper proposes a new GCN model by adding two residual connections (TRC), one is the initial residual connection, and the other is the residual connection to the previous layer. Moreover, a new adjacency matrix and a simpler method for realizing identity mapping are presented and integrated into the TRC-GCN model. The advantage of our TRC-GCN is that the representation of the target node integrates more information, which can learn node embeddings more outstandingly, so as to perform downstream tasks more effectively. Then, we use this model to perform node classification experiments on several datasets (Cora, Citeseer, Pubmed), and found that this model is more excellent than many other models.