Research map / Graph learning / GNN architectures
Accurate and Scalable Estimation of Epistemic Uncertainty for Graph Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Studies GNN confidence-indicator calibration under graph distribution shifts and proposes G-ΔUQ, a single-model uncertainty method extending stochastic anchoring (Δ-UQ) to graphs with input, hidden-layer, and pretrained-readout variants.
Problem
GNN confidence indicators become poorly calibrated under structural, covariate, concept, and graph-size distribution shifts, and architectural expressivity does not fix this; existing ensembles are costly and single-model UQ methods underperform.
Approach
Extends stochastic centering (Δ-UQ anchoring) to graph data by sampling Gaussian-fitted node-feature anchors (input layer), shuffled node-representation anchors at intermediate MPNN layers, or graph-level anchors at readout, enabling partially stochastic and pretrained single-model variants whose epistemic variance modulates predictions.
Key results
- On GOODCMNIST-color concept shift, G-ΔUQ reduces OOD calibration error to 0.3343 vs 0.4392 for vanilla GNN.
- Out-of-distribution calibration error is reduced by stochastic anchoring on 7 of 8 dataset/shift combinations compared with all other methods, per the paper.
- The pretrained G-ΔUQ variant yields the lowest average MAE for generalization gap estimation on GOOD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply G-$\Delta$UQ to GatedGCN | GatedGCN · RotMNIST rotated | ECE: 0.03778 \pm 0.0015 -> 0.0181 \pm 0.00847 | Helped |
| apply G-$\Delta$UQ to GatedGCN | GatedGCN · RotMNIST rotated | accuracy: 0.94716 \pm 0.0022 -> 0.93346 \pm 0.0148 | Hurt |
| use GPS instead of vanilla message-passing GNN | GatedGCN on rotated Superpixel-MNIST · RotMNIST structural distortion shifts | ECE | Hurt |
| add positional encodings (ablation) | GatedGCN on rotated Superpixel-MNIST · RotMNIST structural distortion shifts | ECE | No clear effect |
| increase message-passing depth from 3 to 5 layers (ablation) | GNN on rotated Superpixel-MNIST · RotMNIST structural distortion shifts | ECE | Mixed |
| increase model width (ablation) | GNN on rotated Superpixel-MNIST · 90° and 180° shifts | ECE | Hurt |
| anchor after READOUT rather than at earlier layers (ablation) | G-$\Delta$UQ on DD · DD largest 10% graphs | ECE: 0.5949 \pm 0.018 -> 0.0951 \pm 0.102 | Helped |
| apply G-$\Delta$UQ | GNN on DD, largest 10% graphs · DD largest 10% graphs | ECE: 0.20925 \pm 0.081 -> 0.0951 \pm 0.102 | Helped |
| use G-$\Delta$UQ instead of Deep Ensembles | GIN · GOODCMNIST-color covariate, OOD | calibration error: 0.4954 \pm 0.1773 -> 0.4233 \pm 0.1723 | Helped |
| use G-$\Delta$UQ instead of Deep Ensembles | GIN · GOODMotifLPE-basis covariate, OOD | calibration error: 0.2980 \pm 0.3831 -> 0.3417 \pm 0.2661 | Hurt |
| use pretrained G-$\Delta$UQ instead of end-to-end G-$\Delta$UQ (ablation) | G-$\Delta$UQ · generalization gap prediction on GOOD benchmarks | MAE | Helped |
Surprising findings
- Increasing GNN expressivity via graph transformers (GPS) or positional encodings does not improve calibration under rotation-induced structural shift, and GPS calibrates worse than MPGNN despite similar accuracy.
- Larger width can slightly worsen calibration at high shift levels.
Limitations
- Hyperparameters of G-ΔUQ were not tuned, so accuracy could be improved.
- On GOODMotif-basis, G-ΔUQ has lower OOD AUROC than baselines, attributed to shortcut learning on a simple dataset.
- Evaluated only on graph classification; node classification and link prediction are left to future work.
Experimental setup
Datasets: Superpixel-MNIST rotated MNIST superpixel graphs TUDataset D&D NCI1 NCI109 PROTEINS GOOD benchmark (GOODCMNIST, GOODMotif, GOODSST2)
Benchmarks: Superpixel-MNIST rotated shifts TUDataset size generalization GOOD benchmark
Metrics: accuracy expected calibration error AUROC generalization error prediction MAE
Baselines: Deep Ensembles Monte Carlo dropout temperature scaling SVI vanilla GNN
Builds on: Δ-UQ deep ensembles Monte Carlo dropout GPS graph transformer
Closest papers by meaning
- Energy-based Epistemic Uncertainty for Graph Neural Networks (NeurIPS 2024 · similarity 0.93)
- A Unified Uncertainty Representation for Graph Neural Networks via Doubly-Spectral Stochastic Expansion (NeurIPS 2026 · similarity 0.92)
- Graph Neural Stochastic Diffusion for Estimating Uncertainty in Node Classification (ICML 2024 · similarity 0.91)
- Uncertainty Modeling in Graph Neural Networks via Stochastic Differential Equations (ICLR 2025 · similarity 0.91)
- Uncertainty Estimation for Heterophilic Graphs Through the Lens of Information Theory (ICML 2025 · similarity 0.90)
- Uncertainty Estimation on Graphs with Structure Informed Stochastic Partial Differential Equations (NeurIPS 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.