Do Deep Ensembles Actually Capture Uncertainty in Graph Neural Networks?

2026-05-21 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A new investigation challenges the assumption that deep ensembles effectively capture uncertainty in Graph Neural Networks (GNNs), a belief often extrapolated from their success in computer vision. Benchmarking across seven diverse datasets, the study reveals that standard deep ensembles for message-passing GNNs offer surprisingly minimal improvement over single models. Gains primarily stabilize optimization noise in point predictions, rather than enhancing uncertainty estimates. The research identifies "epistemic collapse," where independently trained networks converge to overly similar predictions, neutralizing the ensemble's ability to capture epistemic uncertainty through disagreement. This collapse is attributed to functional rather than weight-space convexity, meaning distinct parameter solutions yield nearly identical behavior.

Key takeaway

For Machine Learning Engineers and AI Scientists developing GNNs, you should reconsider deep ensembles as the default for uncertainty quantification. This research indicates their limited effectiveness due to "epistemic collapse," where models converge too similarly. Instead, explore alternative uncertainty methods tailored for graph-structured data, or focus on ensemble benefits for stabilizing point predictions rather than robust uncertainty estimates. Your current GNN uncertainty claims might be less reliable than assumed.

Key insights

Deep ensembles for GNNs suffer "epistemic collapse," limiting their uncertainty quantification benefits due to lack of diversity.

Principles

• Deep ensemble success does not seamlessly transfer to graph machine learning.
• Disagreement is fundamental for ensembles to capture epistemic uncertainty.
• Functional convexity can lead to similar predictions despite distinct parameters.

Method

The study investigates standard deep ensembles for message-passing GNNs, benchmarking across seven datasets and employing an aleatoric-epistemic decomposition to analyze uncertainty capture.

In practice

• Re-evaluate deep ensembles for GNN uncertainty quantification.
• Explore alternative methods for GNN uncertainty.
• Focus on optimization stability for GNN point predictions.

Topics

• Deep Ensembles
• Graph Neural Networks
• Uncertainty Quantification
• Epistemic Collapse
• Message Passing Networks
• Machine Learning

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• Epistemic Uncertainty Is Not the Reducible Kind — Standard epistemic uncertainty definitions and measures are inconsistent; reducibility depends on data acquisition methods.
• Semi-Supervised Learning on Graphs using Graph Neural Networks — GNN performance in semi-supervised learning is bounded by approximation, stochastic, and optimization errors, influenced by label scarcity and graph topology.
• Structure-Aware Epistemic Uncertainty Quantification for Neural Operator PDE Surrogates — Restricting uncertainty sampling to the lifting module improves neural operator uncertainty quantification efficiency and accuracy.
• Finite Neural Networks as Mixtures of Gaussian Processes: From Provable Error Bounds to Prior Selection — Finite neural networks can be approximated by Gaussian process mixtures with provable error bounds.
• From Possible to Probable AI Models — AI reliability hinges on distinguishing between what models *can* do and what they *consistently* do.
• Upper Bounds on the Generalization Error of Deep Learning Models via Local Robustness and Stability — A new generalization bound improves accuracy by locally scaling robustness based on stable and unstable samples.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.