Feature augmentation via mutual information in graph neural networks

Abstract

Graph neural networks(GNNs) are becoming the standard for solving a variety of tasks on graph data. Graph data represents the connection information between nodes and node information in the form of node features. In the process of the message-passing mechanism, the hidden nodes representation of each layer is updated from initial node features. The importance of node features in GNNs has been demonstrated in various research, but obtaining complete and accurate node features in real-world cases can be difficult. Thus, there remains a need for an efficient method that can maintain the performance of GNNs in the sparsity of node features. In this paper, we proposed a feature augmentation method based on node similarity. First, to measure node similarity, we define the node mutual information in a graph. Next, based on this node similarity, features are augmented without training another model to estimate node features. We perform experiments on two downstream tasks on five real-world datasets, and the results demonstrate the effectiveness of our method in both link prediction and node classification.