A secure cross-silo collaborative method for imbalanced credit scoring

With the rapid development of information technology, there is an increasing amount of available data that can reflect a borrower’s creditworthiness, providing new avenues for credit scoring innovations. However, such data is commonly distributed across companies in various industries, and how to take advantage of multi-party collaboration while protecting customer data privacy is a major challenge. In this study, we propose an interpretable vertical logistic regression method with adaptive cost sensitivity (IVLR-ACS) for imbalanced credit scoring. Specifically, we construct a collaborative credit scoring method based on vertical logistic regression to preserve the privacy and security of multi-party information. First, to address the imbalanced class distribution problem, we develop an adaptive cost-sensitive (ACS) loss function to enhance the default risk identification of the proposed method. Then, to overcome the potential problem that the proposed method may suffer from malicious attackers adopting other technical means to steal participants’ private information, we design a differential privacy algorithm with adaptive gradient clipping and noise perturbation decay (ADDP) to train the proposed method. Finally, to improve the interpretability of collaborative multi-party credit decision-making, we introduce a feature importance interpretation method inherent to the logistic regression model to analyze the prediction results. We test the performance of the proposed method on eight credit scoring datasets and analyze its interpretability, privacy, complexity, and communication cost. Extensive experimental results demonstrate the competitiveness of the proposed method to utilize multi-party information securely and effectively.