When does distribution shift break graph neural networks calibration?

2026-07-12

AI Summary

This work addresses the lack of theoretical understanding regarding how distribution shifts affect the calibration of graph neural networks (GNNs)—i.e., the alignment between predicted confidence and empirical accuracy—and the reliance of existing methods on target-domain labels. The authors provide the first closed-form theoretical characterization of GNN calibration, revealing that both graph structural changes and feature quality jointly determine whether a model becomes over- or under-confident. They further prove the theoretical optimality of global temperature scaling under this framework. Building on these insights, they propose STAC, a label-free calibration method requiring only source-domain information, and derive an upper bound on expected calibration error under symmetric normalized graph convolution, multi-class classification, and covariate shift assumptions. Experiments demonstrate that STAC significantly improves calibration on synthetic data and validate both the theoretical predictions and the feasibility of unsupervised calibration across five real-world graph datasets.

Abstract

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, and existing graph calibration methods typically rely on labeled validation data from the deployment distribution. In this work, I present the first closed-form theoretical characterization of GNN calibration under distribution shift. I show that calibration is governed by a single scalar quantity that explicitly depends on structural changes between the source and target graphs, as well as feature quality. This characterization precisely identifies when a model becomes over-confident, under-confident, or remains calibrated, and directly yields the optimal temperature scaling strategy. I further extend the analysis to graph convolutional networks with symmetric normalization, multi-class classification, and covariate shift, and derive a theoretical upper bound on the expected calibration error. My analysis also reveals that, under homogeneous distribution shift, a single global temperature is theoretically optimal, providing a principled explanation for why more complex node-wise recalibration methods offer no additional benefit. Building on these theoretical insights, I propose STAC, a source-free, label-free calibration method. Experiments on synthetic benchmarks demonstrate substantial calibration improvements, while evaluations on five real-world graph datasets show that reliable calibration without target labels remains challenging despite the strong predictive power of the theory.

Problem

• distribution shift
• graph neural networks
• model calibration
• expected calibration error
• temperature scaling

Innovation

• graph neural networks
• distribution shift
• model calibration
• temperature scaling
• theoretical analysis