Research map / Graph learning / GNN architectures

Enhancing Trustworthiness of Graph Neural Networks with Rank-Based Conformal Training

AAAI 2025 · Graph learning › GNN architectures · approach: Robustness, privacy, and fairness of GNNs · 4 citations (OpenAlex)

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.

Proposes RCP-GNN, a rank-based conformal prediction method trained end-to-end on GNNs with a differentiable conformity loss to reduce prediction-set size while maintaining target marginal coverage for node classification.

Problem

GNN node classifiers lack rigorous uncertainty estimates; conformal prediction on graphs yields inefficient (large) prediction sets when probabilities are miscalibrated, and standard CP is post-hoc so the model cannot adapt its sets.

Approach

Defines a rank-based conformity score (rank of the true-class probability plus a fractional term) and makes sorting and quantile computation differentiable via sigmoid-based smooth ranking, so a GNN correction model is trained with a conformity loss that penalizes prediction-set size during training while keeping split-CP coverage.

Key results

• RCP-GNN attains target marginal coverage (about 0.950-0.951 at alpha=0.05) on all datasets, the only method besides CF-GNN to do so consistently, with coverage closest to target.
• At alpha=0.1, RCP-GNN gives the smallest prediction sets on all five reported datasets (e.g., Cora 1.18 vs 1.33 for best baseline GATS), with reductions of up to 11.28%.
• Ablation on Cora at alpha=0.05: RCP-THR yields size 1.65 vs RCP-GNN 1.92 on Cora, while RCP-GNN gives 1.26 vs 1.57 for RCP-THR on CiteSeer, showing conformal training matters.

Measured effects reported in the paper

Surprising findings

• CF-GNN does not consistently reduce inefficiency; on Amazon-Photo and Amazon-Computers its prediction sets become larger than the post-hoc baselines.
• On Cora the ablation shows RCP-THR (size 1.65) is smaller than the full RCP-GNN (size 1.92), so the rank-based variant is not uniformly more efficient.

Limitations

• Tuning requires random hyperparameter search over lambda, tau, kappa, layers and learning rate.
• Inefficiency gains are modest on some datasets (e.g., DBLP 1.17 vs 1.18), and ablation results are mixed across datasets.

Experimental setup

Datasets: Cora DBLP CiteSeer PubMed Amazon-Computers Amazon-Photo Coauthor-CS Coauthor-Physics

Benchmarks: Cora DBLP CiteSeer PubMed Amazon-Computers Amazon-Photo Coauthor-CS Coauthor-Physics

Metrics: marginal coverage inefficiency (average prediction set size)

Baselines: Temperature scaling Vector scaling Ensemble temperature scaling CaGCN GATS CF-GNN

Builds on: conformal prediction adaptive prediction sets CF-GNN

Closest papers by meaning

• Valid Conformal Prediction for Dynamic GNNs (ICLR 2025 · similarity 0.88)
• GRAPHLCP: Structure-Aware Localized Conformal Prediction on Graphs (NeurIPS 2026 · similarity 0.87)
• Accurate and Scalable Estimation of Epistemic Uncertainty for Graph Neural Networks (ICLR 2024 · similarity 0.87)
• Similarity-Navigated Conformal Prediction for Graph Neural Networks (NeurIPS 2024 · similarity 0.87)
• A Unified Uncertainty Representation for Graph Neural Networks via Doubly-Spectral Stochastic Expansion (NeurIPS 2026 · similarity 0.87)
• Graph Neural Stochastic Diffusion for Estimating Uncertainty in Node Classification (ICML 2024 · similarity 0.87)