Graph Augmentation for Neural Networks Using Matching-Graphs

Both data access and data collection have become increasingly easy over the past decade, leading to rapid developments in many areas of intelligent information processing. In some cases, however, the amount of data is still not sufficiently large (e.g. in some machine learning applications). Data augmentation is a widely used mechanism to increase the available data in such cases. Current augmentation methods are mostly developed for statistical data and only a small part of these methods is directly applicable to graphs. In a recent research project, a novel encoding of pairwise graph matchings is introduced. The basic idea of this encoding, termed matching-graph, is to formalize the stable cores of pairs of patterns by means of graphs. In the present paper, we propose to use these matching-graphs to augment training sets of graphs in order to stabilize the training process of state-of-the-art graph neural networks. In an experimental evaluation on five graph data sets, we show that this novel augmentation technique is able to significantly improve the classification accuracy of three different neural network models.