基于机器学习的车险欺诈检测方法研究

打开文本图片集

中图分类号：TP181；TP39 文献标识码：A 文章编号：2096-4706（2025）19-0031-08

Abstract：Automatic detection of potential auto insurance fraud can maintain the market order of the auto insurance industry.At present，te traditional MachineLearning methods facethechallengesofsmalldatavolume，unbalancedcategory distributionandfeatureredundancyinimplementingauto insurancefrauddetection.Aimingat theseproblems，this paperselects LightGBM，XGBoost，AdaBoost andRandomForest models，andadopts asetof methods incudingdatapreprocessng，feature selection，hyperparameteroptimizationandcategory weightadjustment.TheexperimentalresultsshowthattheighestROCAUC valuesofeachmodelare0.9069，0.8407，0.8467and0.8513，respectively.At thesametime，throughthecollaborative analysis basedonfeature frequencyanddecisioncontribution，theimportanceofkeyfeatures is discussed，andthe method is aplied to the practical scenarios.Theresearch shows that this method can efectively improve the accuracy and model interpretabilityofauto insurance frauddetection，and provides amore reliableanti-fraudtoolfor the insurance industry.

Keywords：Machine Learning; feature selection;LightGBM; auto insurance fraud detection

0 引言

保险欺诈，尤其是车险欺诈，已成为一个日益严重的社会问题。（剩余12796字）