Robust Meta-Learning over Graphs with Graph Neural Networks

Graph neural networks (GNNs) have well docu-mented success in various learning over graph problems, from drug discovery to recommender systems. These data hungry models however, are challenged in applications with limited available samples. Meta-learning paradigm provides principled methods to empower GNNs to learn rapidly from a few labeled data only, particularly beneficial for tasks with small datasets. Nonetheless, existing meta-learning frameworks for graphs encounter difficulties when the labeled nodes or links used during training are distant from those encountered during testing, rendering meta-learning in graph tasks challenging. In this context, current work leverages a suite of novel techniques to enhance generalization performance of learning over graphs with GNNs. A novel idea here is to further incorporate an extra adversarial update step during “adaptation” phase of conventional meta-learning algorithms. Additionally, our proposed method incorporates curvature information of the loss landscape during meta-updating to facilitate reliable knowledge transfer to downstream tasks with slightly different structures. The proposed methods are simple and with minimal computational costs on training process. Numerical experiments conducted on learning over graph problems validate the superiority of our algorithm over the existing counterparts, showcasing effectiveness when dealing with limited data resources.