基于对比学习与特征交叉融合的贷款违约预测模型研究  

Research on loan default prediction model based on comparative learning and feature cross fusion

作　　者：梁静娟 路新喜 董凌鹤 LIANG Jingjuan;LU Xinxi;DONG Linghe(Beihang University,Beijing 100191,China;Hebei Xingji Talent Resources Development Co.,Ltd.,Shijiazhuang 050011,China)

机构地区：[1]北京航空航天大学,北京100191 [2]河北兴冀人才资源开发有限公司,石家庄050011

出　　处：《计算机应用文摘》2025年第7期69-72,共4页

摘　　要：信贷业务是我国商业银行的核心利润来源,其风险管理水平直接影响银行的盈利能力和金融稳定性。然而,传统的信用评分模型在处理高维稀疏数据、非线性关系以及类别不平衡问题时存在一定的局限性。为此,文章提出了一种基于对比学习与特征交叉融合的贷款违约预测模型(DCN-CL-FL)。在Kaggle信贷违约数据集上的实验结果表明,DCN-CL-FL模型相比次优模型XGBoost,AUC指标提升了2.12%,F1-score提升了4.73%。消融实验进一步验证了各模块的有效性:对比学习模块使AUC提升了4.45%,交叉熵对比损失函数使AUC提升了3.64%。Credit business is the core source of profit for commercial banks in China,and its risk management level directly affects the bank's profitability and financial stability.However,traditional credit scoring models have certain limitations when dealing with high-dimensional sparse data,nonlinear relationships,and class imbalance problems.Therefore,the article proposes a loan default prediction model based on contrastive learning and feature cross fusion(DCN-CL-FL).The experimental results on the Kaggle credit default dataset show that the DCN-CLFL model improved the AUC index by 2.12% and F1 score by 4.73% compared to the suboptimal model XGBoost.The ablation experiment further validated the effectiveness of each module:the contrastive learning module increased AUC by 4.45%,and the cross entropy contrastive loss function increased AUC by 3.64%.

关 键 词：特征交叉 对比学习 贷款违约预测 模型融合 数据不平衡处理 

分 类 号：TP300[自动化与计算机技术—计算机系统结构]