集成学习动态融合的贷款客户失联模式分类与鲁棒优化

打开文本图片集

中图分类号：TP181；F830.49 文献标识码：A 文章编号：2096-4706（2025）17-0110-05

Abstract：Inorder to metthe precise needsoffinancial institutions fortherisk preventionandcontrolofloancustomers' lossofcontact，anensemble learning framework basedondynamic weightfusionisproposed.The framework integrates the advantagesofXGBoost，LightGBMandAdaBoost models，verifiesitsrobustness throughdual stresstestsofoise injection andfeaturelossandanalyzes thedecisionlogicwiththefeatureinterpretabilitymethod.Theexperimentalresultsshowthatthe comprehensiveFl score of the dynamic fusion model is 96.5% ，whichis 1.2% higher than that of the single optimal model.In the noise interference and feature loss scenarios， the F1 score attenuation rates are only 0.83% and 14.79% ，respectively，and thestabilityissignificantlybeterthanthetraditional methodAmongthem，thecontributionofcorefeaturessuchas“overdue amount” and “registered address” to classification decision-making is more than 70% .Through multi-model collaborative optimization，themodelefectively improvestheclasificationaccuacyandanti-interferenceabilityprovides keytechical supportforfancialistitutions tobuildahghlyrobustriskcontrolsystem，andhelpsthedeepintegrationoffinancialdigital transformation and systemic risk prevention and control.

Keywords： uncontactable loan customer; ensemble learning; dynamic weight fusion; robustness evaluation; feature interpretability; financial technology

0 引言

在双循环发展格局与金融数字化转型加速的大背景下，金融行业正经历深刻变革[1。（剩余8449字）