Credit Evaluation of Supply Chain Partners Using Ensemble Learning and Conditional GANs

Abstract

Driven by the deep integration of the industrial internet and digital finance, supply chain risk control is shifting towards dynamic data intelligence. However, extreme sample class imbalances and data feature heterogeneity in real business scenarios severely restrict the accuracy of credit assessment. To overcome the identification blind spots of sparse risk samples, the study proposes an enterprise supply chain partner credit evaluation model combining Conditional Generative Adversarial Networks (CGAN) and Ensemble Learning (EL). First, an improved CGAN reconstructs the true distribution of default samples to alleviate data skew. Then, a Tree-structured Parzen Estimator (TPE) based Bayesian Optimization (BO) algorithm adaptively tunes a two-layer Stacking ensemble framework (including XGBoost, LightGBM, and CatBoost) for deep heterogeneous feature mining. Results show that the model exhibits excellent comprehensive classification performance, with an F1 score of 0.949 and a Kolmokolov-Smirnov statistic of 0.62, significantly outperforming benchmark models such as Attentive Interpretable Tabular Learning (TabNet). In credit simulation, the model achieves a 94.5% default rejection rate while maintaining high credit coverage, avoiding huge bad debt losses. The study effectively solves the problem of capturing minute risks in high-dimensional heterogeneous spaces, providing a scientific basis with high economic value for financial institutions to optimize credit resource allocation and reduce decision-making uncertainty.