Comparison of various classification algorithms for credit card fraud detection

This paper thoroughly compares several classification algorithms and provides a methodical investigation of credit card fraud detection. This research departs from traditional methods by taking into account both mainstream and unorthodox algorithms in response to the increasing sophistication of fraudulent actions. Our work attempts to offer nuanced insights into the relative efficacy of these algorithms using a combination of theoretical analysis and empirical evaluations conducted in the actual world. In the continuous fight against financial misconduct, this research helps to develop more accurate and flexible fraud detection systems by questioning conventional wisdom and encouraging creativity. We conducted our analysis using the traditional European Credit Card Fraud Detection dataset, which is accessible on Kaggle. Many traditional approaches of categorization, such as Logistic Regression on the dataset, Support Vector Machine and KNN were used, and the results showed extremely high classification accuracy (>85% ROC-AUC Score). Then, using ensemble models such as Random Forest, Gradient Boosting, XGBoost, AdaBoost, and CatBoost, the task of classifying fraud based on transaction data was completed with a higher accuracy (>91.1% AUC-ROC Score). The Hyperparameter Tuning work was performed using the Randomized Search CV approach. We used Explainable AI (XAI) techniques to make sense of the outcomes.