Contra2: A one-step active learning method for imbalanced graphs

Abstract

Graph active learning (GAL) is an important research direction in graph neural networks (GNNs) that aims to select the most valuable nodes for labeling to train GNNs. Previous works in GAL have primarily focused on the overall performance of GNNs, overlooking the balance among different classes. However, graphs in real-world applications are often imbalanced, which leads GAL methods to select class-imbalanced training sets, resulting in biased GNN models. Furthermore, due to the high cost of multi-turn queries, there is an increasing demand for one-step GAL methods, where the entire training set is queried at once. These realities prompt us to investigate the problem of one-step active learning on imbalanced graphs.

In this paper, we propose a theory-driven method called Contrast & Contract (Contra²) to tackle the above issues. The key idea of Contra² is that intra-class edges within the majority are dominant in the edge set, so contracting these edges will reduce the imbalance ratio. Specifically, Contra² first learns node representations by graph contrastive learning (GCL), then stochastically contracts the edges that connect nodes with similar embeddings. We theoretically show that Contra² reduces the imbalance ratio with high probability. By leveraging a more evenly distributed graph, we can achieve a balanced selection of labeled nodes without requiring any seed labels. The effectiveness of Contra² is evaluated against various baselines on 11 datasets with different budgets. Contra² demonstrates remarkable performance, achieving either higher or on-par performance with only half of the annotation budget on some datasets.