保险反欺诈预测  

Insurance Anti-Fraud Forecasting

作　　者：李欣朵 余旺 

机构地区：[1]重庆理工大学理学院,重庆

出　　处：《国际会计前沿》2024年第6期964-975,共12页Frontiers of International Accounting

摘　　要：保险行业不断发展,保险欺诈现象在社会上越来越常见,不仅给保险公司带来巨大的损失,而且对社会发展产生了极其不利的影响。因此需要建立保险欺诈预测模型,从一定程度上遏制保险欺诈行为。利用700条保险反欺诈数据,其中特征有39个,标签为是否保险欺诈。首先,对数据进行缺失值、重复值、日期数据转化、类别数据离散化等数据预处理,以便于后续模型的建立。同时考虑到特征之间可能存在共线性问题,对特征进行方差分析、相关性分析筛除掉特征之间相关性大的多余特征,以及与标签相关性小的无用特征。对上述数据处理和特征筛选后的数据建立机器学习模型,来预测保险欺诈行为。选择使用用于分类预测的常见模型:logistics回归、knn以及Bagging集成学习随机森林模型和Boosting集成学习的LightGBM。从各模型对测试集的预测结果评估可以发现LightGBM模型的整体的模型预测性能最好,预测准确率达到88%,可以作为保险欺诈预测的参考模型。而logistics回归、knn模型存在对保险诈骗的预测准确率较低,将大部分数据预测为非保险诈骗数据,因此实际应用性较差。With the continuous development of the insurance industry, insurance fraud has become more and more frequent in society, which not only brings huge losses to insurance companies, but also has an extremely adverse impact on social development. Therefore, it is necessary to establish an insurance fraud prediction model to curb insurance fraud to a certain extent. 700 insurance anti-fraud data are used, of which there are 39 features and the label is whether it is insurance fraud. First, the data is preprocessed by missing values, duplicate values, date data conversion, and category data discretization to facilitate the establishment of subsequent models. At the same time, considering the possible collinearity problem between features, variance analysis and correlation analysis a

关 键 词：保险欺诈 方差分析 相关性分析 机器学习预测 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]