Abstract: Solving classification problems using machine learning methods, data scientists often face the problem of data imbalances. Class imbalance is common in financial sector, in particular for the task of fraud detection in car insurance. Training models on unbalanced data can lead to misclassifications and large numbers of false positives due to the tendency of the model to classify observed cases as the majority class.

This paper deals with the study of ways to solve the problem of class imbalance in the task of insurance claims classifying. To solve this problem, a database in the field of auto insurance was used, which provide information about the presence or absence of fraudulent customer claims. The class of fraudulent cases that interests us the most is represented in the database by three times fewer records than for legitimate claims. Oversampling techniques including random oversampling and SMOTE were applied to avoid modeling problems on unbalanced data. Evaluation of the results obtained on different samples indicates that balancing methods can significantly improve the quality of the classification.

Logistic regression, support vector machine, k-nearest neighbors classifier, Bayesian classifier, decision tree, random forest and perceptron type neural network were built on the obtained datasets. A comparative analysis of the models’ qualities allowed to determine the best methods for detecting fraudulent claims. For both datasets, logistic regression and neural network were recognized as such methods, having a high level of fraud detection combined with a good predictive power of the model.