融合机器学习算法的银行客户信用风险评估研究

Research on Credit Risk Assessment of Bank Customers Based on Integrating Machine Learning Algorithm

摘要: 在金融数字化转型加速的背景下，银行客户信用风险决策面临“样本不均衡”与“误判成本高昂”的双重挑战，亟需兼顾整体精度与少数类召回的稳健模型。本文基于阿里云天池22,500名银行客户信贷数据，系统对比逻辑回归、随机森林与XGBoost三种机器学习模型，发现单一模型在召回率与F1指标上的不足；进一步提出融合模型框架，结合软投票加权平均与代价敏感学习，在不改变数据分布的前提下放大正类梯度权重，有效提升召回率至80.17%，较最优单一模型提高4.07%，AUC达0.8913，准确率保持在83.13%，为银行评估高风险客户提供了可解释、可落地的技术路径。

Abstract: In the context of the accelerating digital transformation in finance, bank customer credit risk decision-making faces the dual challenges of “sample imbalance” and “high misjudgment costs”. This situation calls for a robust model that balances overall precision and minority class recall. This study systematically compares three machine learning models—Logistic Regression, Random Forest, and XGBoost—based on the credit data of 22,500 bank customers from Alibaba Cloud Tianchi, finding that a single model falls short in recall rate and F1 metrics. It further proposes an ensemble model framework that integrates soft voting weighted averaging and cost-sensitive learning, effectively increasing the positive class gradient weight without altering the data distribution, thereby boosting the recall rate to 80.17%, an improvement of 4.07% over the optimal single model, with an AUC of 0.8913 and an accuracy maintained at 83.13%, providing an interpretable and actionable technical pathway for banks to assess high-risk customers.