CREDIT CARD FRAUD DETECTION USING ML

Abstract

This thesis investigates the potency of supervised machine learning techniques for detecting credit card fraud in highly imbalanced transaction data. Using a publicly available dataset of over 1.6 million transactions only 0.5% of which are fraudulent five approaches (Random Forest, Decision Tree, Naive Bayes, Logistic Regression, and LSTM) were carried out and evaluated on a 70:30 train–test split. To confront the hurdle of skewed class ratio, the Synthetic Minority Over-sampling Technique (SMOTE) was employed in coordination with Random Forest and LSTM, generating realistic synthetic fraud instances. Model performance was assessed via confusion matrices and assessment criterion accuracy, precision, sensitivity, F1-score, and ROC-AUC alongside computational efficiency. On the original imbalanced data, Random Forest achieved the highest accuracy (99.77%) but exhibited low recall, indicating many missed fraud cases. After SMOTE, all models showed marked improvement in recall and F1-score, with LSTM outperforming others (99.87% accuracy, 93.73% recall, 92.85% F1-score, 99.75% ROC-AUC). These findings demonstrate that combining deep learning with targeted oversampling yields the most balanced fraud detection performance. The study offers feasible guidance targeted at financial entities pursuing adaptive, data-driven fraud prevention solutions and arranges the groundwork for future research into real‐ time and hybrid detection systems.