Detection of Fraud in Credit Card Transactions (Academic research)

Summary

• The study uses Kaggle's public Credit Card Fraud Detection dataset of 284,807 transactions, only 492 of which are fraudulent.p.1
• Fraud represents approximately 0.172% of all transactions, illustrating extreme class imbalance.p.5
• Data were split 80%-20% into training and test sets, with 5-fold stratified cross-validation on the training data.p.7
• Logistic regression has a ROC AUC of 0.98 but a PR AUC of only 0.68.p.10
• XGBoost, the selected model, reaches an AUC ROC of 0.97, an AUC PR of 0.81 and an F1 score of 0.871.p.13
• Naive Bayes has the lowest PR-AUC (0.15) and SVC achieves 0.34, despite high ROC-AUC values.p.15

Key insights

1. Each transaction is described by 30 variables, of which V1 through V28 come from a PCA transformation.p.5
2. SMOTE was applied only to the training subsets of each fold to prevent data leakage.p.7
3. KNN achieves an ROC AUC of 0.90 and a PR AUC of 0.76.p.11
4. The decision tree records an ROC AUC of 0.86 and a PR AUC of 0.56.p.11
5. Random Forest achieves an ROC AUC of 0.94 and a PR AUC of 0.81.p.12
6. SVC has a recall of 0.72 and a precision of approximately 0.24, giving an F1-score of approximately 0.36.p.17
7. A decision threshold of 0.5 was used for all classifiers.p.9
8. The proposed XGBoost model with stratified cross-validation reports an accuracy of 0.91 in the literature comparison.p.18