Variational Bayesian Graph Neural Networks for Uncertainty-Aware Credit Evaluation

Abstract

Enterprise credit evaluation in international trade faces critical challenges including data heterogeneity, dynamic industry risks, and inadequate uncertainty quantification. This study proposes a Variational Bayesian Graph Neural Network (VB-GNN) framework to address these issues by integrating variational Bayesian (VB) inference with graph neural networks (GNNs). The VB-GNN model quantifies uncertainty in credit assessment through posterior distribution estimation of node embeddings, optimized via Kullback-Leibler (KL) divergence. It incorporates industry-adaptive prior distributions (e.g., weighted inventory turnover for manufacturing and R&D intensity for technology sectors) and dynamic threshold adjustment to adapt to sector-specific risk characteristics. Experimental results on 100,000 enterprises across 8 industries show that VB-GNN reduces entity resolution errors by 37% and improves default prediction accuracy by 28% compared to traditional models. The framework outputs credit scores with confidence intervals (e.g., 720±35), satisfying the risk transparency requirements of ISO 31000. This work advances uncertainty-aware credit evaluation by enabling robust risk communication and cross-industry adaptability.