arxivcs.LG2026-07-12
When does distribution shift break graph neural networks calibration?
Graph neural networks (GNNs) are increasingly deployed in real-world applications where distribution shift is un-avoidable. However, how such shifts affect model calibration, defined as the agreement between predictive confidence and actual accuracy, remains poorly understood, an…