Evidential Uncertainty Probes for Graph Neural Networks

2025-03-11

AI Summary

In high-stakes graph learning tasks—such as drug discovery and financial fraud detection—pretrained graph neural networks (GNNs) lack plug-and-play capability for quantifying both epistemic and aleatoric uncertainty. To address this, we propose EPN, an evidence probing framework that requires no fine-tuning or retraining: it extracts Dirichlet evidence from pretrained GNN representations via a lightweight MLP probe head. We further introduce EPN-reg, the first evidence-based regularization method with theoretical guarantees for calibrated epistemic uncertainty estimation. EPN is architecture-agnostic and compatible with any pretrained GNN. Empirically, it achieves state-of-the-art performance in uncertainty calibration and out-of-distribution detection across multiple benchmarks, with zero additional training overhead. By enabling reliable, interpretable uncertainty quantification without modifying pretrained models, EPN significantly enhances the trustworthiness of GNN deployment in safety-critical applications.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Graphical Models
• Data Mining & Knowledge Management: Representing, Reasoning, and Using Provenance, Trust

Application Category

• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Web Mining and Content Analysis: Large pretrained models with web data
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Accurate quantification of both aleatoric and epistemic uncertainties is essential when deploying Graph Neural Networks (GNNs) in high-stakes applications such as drug discovery and financial fraud detection, where reliable predictions are critical. Although Evidential Deep Learning (EDL) efficiently quantifies uncertainty using a Dirichlet distribution over predictive probabilities, existing EDL-based GNN (EGNN) models require modifications to the network architecture and retraining, failing to take advantage of pre-trained models. We propose a plug-and-play framework for uncertainty quantification in GNNs that works with pre-trained models without the need for retraining. Our Evidential Probing Network (EPN) uses a lightweight Multi-Layer-Perceptron (MLP) head to extract evidence from learned representations, allowing efficient integration with various GNN architectures. We further introduce evidence-based regularization techniques, referred to as EPN-reg, to enhance the estimation of epistemic uncertainty with theoretical justifications. Extensive experiments demonstrate that the proposed EPN-reg achieves state-of-the-art performance in accurate and efficient uncertainty quantification, making it suitable for real-world deployment.

Problem

• Quantify uncertainty in Graph Neural Networks for high-stakes applications.
• Enable uncertainty estimation without retraining pre-trained GNN models.
• Enhance epistemic uncertainty estimation using evidence-based regularization techniques.

Innovation

• Plug-and-play framework for GNN uncertainty quantification
• Lightweight MLP head extracts evidence from representations
• Evidence-based regularization enhances epistemic uncertainty estimation