Structure and Features Fusion with Evidential Graph Convolutional Neural Network for Node Classification

Anonymous

Abstract: Recently, text-enhanced network representation learning has achieved great success by taking advantage of rich text information and network structure information. However, content-rich network representation learning and quantifying classification uncertainty are challenging when it comes to integrating complex structural dependencies and rich content features at an evidence level. In this paper, we propose an evidential graph representation learning model (EGCN), which can not only fuse network structure and content information into a more complete and powerful representation for each node, but also assess the quality of graph node features to improve classification accuracy. To achieve better fusion, we integrate the node's features representation into structure-aware representation through a delivery operator. Besides, to overcome the difficulty of predicting node classification confidence, we employ a novel module based on Dirichlet distribution theory of evidence and subject opinion learning to collect the evidence of the class probabilities. Experimental results on three real-world networks show that our model can improve both node classification accuracy and robustness as compared to all baselines.