Enhancing Fraud Detection with Gradient Boosting Algorithms in Credit Card Transactions

The paper shows the trade-offs between interpretability, computation cost and accuracy of many algorithms for fraud detection from machine learning perspective which provides important clues in this field and potential improvement paths. Linear model logistic regression is pitted against the gradient-boosting methods CatBoost and LightGBM, with excellent ability to handle complex connections in the data. To assess how well the models are performing, confusion matrices and visualizations such as ROC curves, accuracy comparisons are performed. While Logistic Regression results are respectable, as can be noticed from the higher AUC-ROC values and better F1 scores, gradient-boosting methods like CatBoost, LightGBM have a slight edge when it comes to predictive performance. The paper further mentions the trade-offs between performance and simplicity in its conclusion - including gradient-boosting models' usability for running on complex datasets like those present in credit card fraud detection use-cases. The results showed that while Logistic Regression performs reasonably on basic data, two gradient-boosting models, CatBoost and LightGBM, performed better overall, as evidenced by higher AUC-ROC values and F1 score.