CatBoost and Bayesian Optimization Algorithm-Based Classification of Fraudulently Insured Persons and Random Forest-Based Fraud Prediction Combining Multiple Attention Mechanisms

With the development of science and technology, machine learning is gradually applied to the field of insurance to cope with the complex insurance fraud that is difficult to identify by traditional means. This paper is based on motor insurance fraud data, combined with the basic information of the insured, descriptive statistics and variable correlation analysis, screening out the top ten factors that have a greater impact on insurance fraud. By combining various machine learning algorithms, the fraudulent insurance situation is classified and judged as well as predictive modeling, and the model is optimized to improve the accuracy. It is found that accident severity has the strongest correlation with insurance fraud, followed by contact organization, hobbies, accident type and insurance liability cap. The Catboost algorithm with Bayesian optimization was finally used to achieve a prediction accuracy of 0.95. At the same time, the constructed random forest and multi-attention mechanism model can output the prediction probability and average attention weight map of each case, which shows that the model pays more attention to the severity of the accident. The results of the study provide effective technical support for the identification and prevention of insurance fraud, and propose the construction of a risk scoring mechanism to assist in the development of personalized programs.