Improved Representations using Augmented Graphs

Gagan Raj Gupta, Soumajit Pramanik, Milind Kesar Thummala, Anirban Haldar

Abstract: Node classification has become an essential problem with several real-world applications. In the last few years, graph neural networks have been widely used for classifying nodes on large networks. However, the success of existing methods is found to be limited while classifying low-degree nodes. In this paper, we propose a suite of neighborhood augmentation strategies that enhance the general classification accuracy of graph neural networks as well as improve the results for the aforementioned nodes. On top of the structural neighborhood, here we incorporate the notion of representation space neighborhood for further improving node representations. In particular, with the help of locality-sensitive hashing, we propose a scalable method to augment nodes with their representation space neighbors. Our LSH-Stack formulation resulted in 6.6% relative improvement in Micro F1 scores on average in graph benchmarks over GraphSAGE GNN. Finally, we discuss how our proposed neighborhood augmentation strategy is generic and can be used to improve the quality of the learned representation in other domains such as images, where there is no graph structure to begin with.