Dual Channel Graph Convolutional Networks via Personalized PageRank

Dear Editor, This letter presents a node feature similarity preserving graph convolutional framework PD2G. Graph neural networks (GNNs) have garnered significant attention for their efficacy in learning graph representations across diverse real-world applications. They adeptly explore graph structure and node features, employing aggregation and transformation mechanisms within neighborhoods. However, some studies have indicated a trend among GNN models, which prioritize structural information over node feature information. Unfortunately, in numerous scenarios, the significance of node features remains paramount. On top of that, it is acknowledged that datasets frequently encompass noisy structural and feature components, which may not be appropriately refined for subsequent analysis. Inspired by these, we propose a novel GNN framework PD2G, a personalized PageRank (PPR) driven dual GNN, to capitalize on the rich node feature information. Specifically, our approach includes the integration of consistency loss within the feature channel to feature biases in the model. Besides, we leverage the cross-channel alignment strategy to balance information from graph structure and node features. Extensive experiments show the superiority of our solutions when contrasted with baselines.