Graph Neural Network-based Node Classification with Hard Sample Strategy

Existing graph neural networks (GNNs) usually use a balanced class distribution to learn node embeddings over graph data. When dealing with an imbalanced class distribution, they tend to bias to nodes in majority classes, while nodes from minority classes are under-represented. To meet this challenge, this paper introduces an effective GNN-based node classification model with Hard Sample Strategy (GNN-HSS) to handle class-imbalanced graph data. The proposed GNN-HSS model first uses a two-layer graph convolutional network (GCN) to get node embeddings, and then performs a clustering analysis procedure to make the node embeddings more representative and easier to classify. In particular, a hard sample strategy is given to ensure that the embeddings of hard nodes are correctly represented. The experiments show that GNN-HSS outperforms state-of-the-art methods in node classification tasks.