AdaCal-GNN: Adaptive Confidence Calibration for Graph Neural Networks

Abstract

Graph Neural Networks (GNNs) often produce poorly calibrated predictions, where the predicted probabilities do not accurately reflect the true likelihood of correctness. Recent studies indicate that miscalibration in GNNs exhibits a heterogeneous pattern: while neighborhood aggregation induces an overall underconfidence bias, certain nodes with specific local topologies can be overconfident. Existing calibration methods typically rely on global post-hoc adjustments or regularization, which are insufficient to handle node-level variability. In this work, we introduce AdaCal-GNN, an end-to-end adaptive calibration framework. Leveraging the observation that controlling the L2 norm of pre-softmax logits is equivalent to node-specific temperature scaling, AdaCal-GNN predicts optimal logit norms directly from hidden representations. This approach dynamically enhances the confidence of underconfident nodes while tempering overconfident ones during training. Experiments on benchmark datasets (Cora, Citeseer, PubMed, and CoraFull) show that AdaCal-GNN substantially reduces Expected Calibration Error (ECE) without sacrificing classification accuracy, offering a reliable and generalizable calibration solution for GNNs.