Graph neural networks (GNNs) have generalized deep learning methods into graph-structured data with promising performance on graph mining tasks. However, existing GNNs often meet complex graph structures with scarce labeled nodes and suffer from the limitations of non-robustness, over-smoothing, and overfitting. To address these issues, we propose a simple yet effective GNN framework---Graph Random Neural Network (Grand). Different from the deterministic propagation in existing GNNs, Grand adopts a random propagation strategy to enhance model robustness. This strategy also naturally enables Grand to decouple the propagation from feature transformation, reducing the risks of over-smoothing and overfitting. Moreover, random propagation acts as an efficient method for graph data augmentation. Based on this, we propose the consistency regularization for Grand by leveraging the distributional consistency of unlabeled nodes in multiple augmentations, improving the generalization capacity of the model. Extensive experiments on graph benchmark datasets suggest that Grand significantly outperforms state-of-the-art GNN baselines on semi-supervised graph learning tasks. Finally, we show that Grand mitigates the issues of over-smoothing and overfitting, and its performance is married with robustness.

提出了一个名为GRAPH RANDOM NEURAL NETWORKS (GRAND)的图随机神经网络框架，通过设计随机传播策略来进行图数据增强，并利用一致性正则化优化不同数据增强下未标记节点的预测一致性，实验证明该方法显著优于现有的基准GNN在半监督节点分类上，并且缓解了过度平滑和非鲁棒性问题，展现了比现有GNN更好的泛化行为。

图随机神经网络用于图半监督学习