Regularizing Graph Neural Networks via Consistency-Diversity Graph Augmentations

Despite the remarkable performance of graph neural networks (GNNs) in semi-supervised learning, it is criticized for not making full use unlabeled data and suffering from over-fitting. Recently, augmentation, used to improve both accuracy generalization GNNs, has received considerable attentions. However, one fundamental question how evaluate quality augmentations principle? In this paper, we propose two metrics, Consistency Diversity, aspects augmentation correctness generalization. Moreover, discover that existing fall into a dilemma between these metrics. Can find satisfying consistency diversity? A well-informed answer can help us understand mechanism behind GNNs. To tackle challenge, analyze representative learning algorithms: label propagation (LP) regularization (CR). We LP utilizes prior knowledge graphs CR adopts variable promote diversity. Based on discovery, treat neighbors as capture embodying homophily assumption, which promises high augmentations. further diversity, randomly replace immediate each node with its remote neighbors. After that, neighbor-constrained proposed enforce predictions augmented be consistent other. Extensive experiments five real-world validate superiority our method improving