Learning Graph Neural Networks on Feature-Missing Graphs

Jun Hu, Jinyan Wang, Quanmin Wei, Du Kai, Xianxian Li

Abstract: Graph neural networks have demonstrated state-of-the-art performance in many graph analysis tasks. However, relying on both node features and topology completeness can be challenging, especially as node features may be completely missing. Existing efforts that direct node feature completion suffer from several limitations on feature-missing graphs. In this paper, we propose a novel and general extension for running graph neural networks on feature-missing graphs via complete missing node feature information in the embedding space, called GNN-FIC. Specifically, it utilizes a Feature Information Generator to simulate missing feature information in the embedding space and then completes the node embedding using the predicted missing feature information. Additionally, GNN-FIC introduces two alignment mechanisms and a relation constraint mechanism, aiding in generating high-quality missing feature information. Extensive experiments on four benchmark datasets have shown that our proposed method provides consistent performance gains compared with several advanced methods.