Node Augmentation Methods for Graph Neural Network based Object Classification

Graph neural networks (GNNs) are powerful models for learning representations of relational data, and have achieved impressive results both academically and industrially. To further enhance the performance of GNNs on the most studied node classification problem, we present NodeAug, a novel augmentation method that operates on graph-structured data, yielding virtual nodes by mixing pairs of nodes and corresponding graph structures. We first generate isolated virtual nodes with features and labels being convex combinations of existing nodes, without considering edges, which is termed as NodeAug-I and can enhance many GNN variants, demonstrating the effectiveness of node augmentation. Still, it does not exploit the graph structure, which is essential in GNNs while incorporating the structure in augmentation is a key challenge. Regarding this, we further propose two novel algorithms that can mix information of neighbors as well, taking graph structures into account. We conduct experiments on a wide range of benchmarks, including Cora, CiteSeer, Pubmed, CoraFull, Amazon and CLUSTER, and obtain considerable enhancement for well-known models, including GCN, GraphSAGE, GIN and GAT.