APPLICATION OF GRADIENT BOOSTING ALGORITHMS FOR DETECTING FINANCIAL FRAUD

Kurtash Nikita Sergeevich, Bryutova Sofia Danilovna, Molodyakov Sergey Alexandrovich

The current problem of automatic detection of fraudulent transactions in the banking sector, complicated by a strong imbalance in data classes, is considered. The evolution of boosting methods from the classic AdaBoost to the modern XGBoost is presented. A comparative analysis of the basic and optimized model configurations was performed on a combined dataset of 300,000 transactions with a 5% fraud rate. Hyperparameter optimization was performed using a Bayesian approach (the TPE algorithm of the Optuna framework). The results showed an increase in F1-score by 57.8% and Precision by 135.6%, indicating a significant reduction in false positives. The dominant influence of tree depth and learning rate on classification quality was revealed. It was demonstrated that proper gradient boosting tuning allows for effective class separation without the use of additional data balancing methods.

Keywords:machine learning, gradient boosting, AdaBoost, XGBoost, fraud detection, unbalanced data, hyperparameter optimization.