Research map / Graph learning / GNN architectures
Moderate Message Passing Improves Calibration: A Universal Way to Mitigate Confidence Bias in 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.
Shows that GNN confidence bias depends on message aggregation (deep/high-label-rate GNNs are overconfident, shallow ones underconfident) and proposes AU-LS, an adaptive unified label smoothing loss that uses positive or negative smoothing to calibrate GCN and GAT.
Problem
GNN confidence calibration is biased in both directions: prior work reports underconfidence, but deeper GNNs and high label rates can be overconfident. The paper asks how message passing relates to this bias and how to correct both directions.
Approach
Links confidence bias to the amount of message aggregation and proposes AU-LS, which unifies positive label smoothing (for overconfidence) and negative label smoothing (for underconfidence) into one loss with an epoch-dependent harmonic-series adaptive parameter.
Key results
- On 2-layer GCN and GAT, AU-LS attains the lowest ECE in almost all cases, e.g., Cora GCN-2 L/C=20 ECE 0.0356 vs 0.1347 uncalibrated.
- For 3-layer GNNs, TS, MS and CaGCN often fail to improve calibration, while AU-LS reduces ECE (e.g., Cora GCN-3 L/C=40: 0.0252 vs 0.0369 uncalibrated).
- Ablation shows adding the unified and adaptive components progressively lowers ECE relative to plain label smoothing.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply AU-LS calibration | GCN (2-layer) · Cora, L/C=20 | ECE (M=20): 0.0401 -> 0.0356 | Helped |
| apply AU-LS calibration | GCN (2-layer) · Citeseer, L/C=20 | ECE (M=20): 0.0595 -> 0.0435 | Helped |
| apply AU-LS calibration | GCN (2-layer) · Pubmed, L/C=20 | ECE (M=20): 0.0405 -> 0.0376 | Helped |
| apply AU-LS calibration | GCN (3-layer) · CoraFull, L/C=20 | ECE (M=20): 0.1221 -> 0.0577 | Helped |
| add unified smoothing to label smoothing (ablation) | GCN-2 with LS · Cora, L/C=20 | ECE (M=20): % -> 0.0715 | Helped |
| add the adaptive module to unified label smoothing (ablation) | GCN-2 with ULS · Cora, L/C=20 | ECE (M=20): 0.0715 -> 0.0356 | Helped |
| add the adaptive module to unified label smoothing (ablation) | GCN-3 with ULS · Cora, L/C=20 | ECE (M=20): 0.0449 -> 0.0297 | Helped |
| use positive-ϵ label smoothing (ablation) | underconfident GCN · GCN calibration scenarios in Table 3 | ECE (M=20) | Hurt |
Surprising findings
- Plain positive label smoothing worsens calibration for underconfident 2-layer GCNs, and 3-layer GNNs can be overconfident, contradicting the prevailing view that GNNs are always underconfident.
- Post-hoc methods (TS, MS, CaGCN) fail to improve well-calibrated 3-layer GNNs, whereas AU-LS still helps.
Limitations
- Theoretical propositions are argued under simplifying assumptions (mean aggregation, balanced classes) and are heuristic.
- Experiments are limited to GCN and GAT on four citation datasets with small label rates.
- Hyperparameters epsilon and gamma are tuned per dataset and label rate to get optimal performance.
Experimental setup
Datasets: Cora Citeseer Pubmed CoraFull
Benchmarks: Cora Citeseer Pubmed CoraFull
Metrics: ECE (M=20)
Baselines: Uncalibrated Temperature Scaling Matrix Scaling CaGCN Label Smoothing Unified Label Smoothing
Builds on: label smoothing temperature scaling CaGCN
Closest papers by meaning
- Understanding and Mitigating Under-Confidence in GNNs from the Final Layer (NeurIPS 2026 · similarity 0.92)
- Improving GNN Calibration with Discriminative Ability: Insights and Strategies (AAAI 2024 · similarity 0.88)
- Accurate and Scalable Estimation of Epistemic Uncertainty for Graph Neural Networks (ICLR 2024 · similarity 0.87)
- Robust Graph Neural Networks via Unbiased Aggregation (NeurIPS 2024 · similarity 0.86)
- Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence (ICML 2025 · similarity 0.86)
- The Effectiveness of Curvature-Based Rewiring and the Role of Hyperparameters in GNNs Revisited (ICLR 2025 · similarity 0.86)
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.