Enhancing Graph Neural Networks by Structure-wise Graph Data Augmentation

Graph neural network is a new neural network model in recent years, whose advantage lies in processing graph structure data. In the era of big data, people can collect a large amount of graph data for training neural network models, but in other scenarios, they can only use limited, missing, and noisy data. Such data will undoubtedly bring negative benefits to the performance of the neural network model. In the field of image processing, there are already many mature methods to deal with noisy pictures or small sample data. However, related research on graph data still needs to be carried out. This paper introduces a graph data augmentation technique based on edge probability prediction. Assume that the collected graph data has edge noise, which means we cannot confirm whether there exists an edge between two nodes that are recorded not connected in the adjacency matrix. The graph autoencoder to represent the probability that there is an edge between each node in the graph. By adding those edges with high probability that do not exist in the adjacency matrix to the graph, the connectivity of the same kind of nodes can be enhanced, thereby improving the final accuracy of the model. In this paper, experiments are carried out on the reference network datasets of papers such as cora, and classic graph neural networks such as gcn and gat are selected as the baseline for comparison. Results shows a significant improvement is achieved.