Contrastive Graph Convolutional Networks with adaptive augmentation for text classification

Yintao Yang, Rui Miao, Yili Wang, Xin Wang

Abstract: Highlights•We study a novel problem of applying supervised graph contrastive learning to text classification, and propose a contrastive graph representation learning framework called CGA2TC.•The CGA2TC not only makes full use of labeled and unlabeled data but also randomly utilizes some nodes in the contrastive training process instead of all nodes to reduce resource consumption.•The CGA2TC with adaptive augmentation enables more effective preservation of the graph’s structure and obtains robust text representations for the text classification task.•We select two centrality measures for nodes in the adaptive augmentation section to measure the importance of the nodes, and then remove the edges of the nodes based on their importance to produce two contrastive views.•We skillfully divide the GNN model into two parts, namely, a feature extractor and a classifier to better match the contrastive learning framework.