Multi-strategy adaptive data augmentation for Graph Neural Networks

Abstract

While existing Graph Neural Networks (GNNs) have demonstrated exceptional performance in semi-supervised learning, they are criticized for not fully utilizing unlabeled data and graph structure, leading to over-fitting. Recently, data augmentation as a technique gaining significant attention recently for enhancing the generalization capabilities of Graph Neural Networks (GNNs), is regarded as a promising solution to mitigate the issue of over-fitting. Nonetheless, existing studies focus on designing a simple augmentation strategy tailored to a specific dataset, without ensuring a diverse and consistent distribution of augmented graphs during the training process. Moreover, formulating different graph augmentation strategies for diverse datasets relies heavily on domain knowledge due to the distinct distributions inherent in various graphs. Designing a universal augmentation strategy which can adapt to different datasets remains a challenging endeavor. To resolve the thorny issue, we introduce an adaptive augmentation framework called MSA-AUG, which can adaptively integrate multiple heuristic strategies for GNNs. Specifically, in order to enhance the representation quality of GNNs, we first take advantage of various augmentation strategies from global, local and label perspectives. Then we integrate them into a set of candidate strategies, which is model-agnostic for any GNNs. Then, we construct a search space of candidate strategies and take each candidate-strategy with different selection weights and operation magnitudes. The learned parameters are leveraged to explore different selection weights and operation magnitudes for ensuring that the contribution of various graph augmentation strategies is well-adjusted. Finally, we employ a search algorithm based on density matching to dynamically explore the candidate strategies and take the best augmentation strategy on various datasets with different distributions for improving the generalization ability of GNNs. Extensive experiments on diverse benchmark datasets demonstrate that MSA-AUG can improve the performance of GNNs with varying structures of backbones.