Credit card fraud detection under severe class imbalance
Binary classification pipeline addressing extreme class imbalance in electronic transaction fraud detection.
- Model
- Logistic regression baseline
- Problem
- Binary classification, positives well under 1%
- Handling
- Resampling to correct class balance
- Evaluated
- Precision, recall, AUC-ROC
Problem Formulation
In electronic credit card transaction data, legitimate transactions vastly outnumber fraudulent activity, with positive cases accounting for less than 0.2% of total events. Under this level of class skew, raw classification accuracy is uninformative, as a trivial majority-class classifier achieves >99.8% accuracy while detecting zero fraud.
Methodology & Resampling
A supervised classification pipeline was developed in Python using scikit-learn. Training distributions were adjusted using Synthetic Minority Over-sampling (SMOTE) and targeted undersampling strategies to improve minority class representation without degrading classifier calibration on validation sets.
Threshold Optimization & Evaluation
Models were evaluated using Precision-Recall Area Under the Curve (PR-AUC), F1-score, and cost-matrix analysis. Decision thresholds were tuned explicitly to balance the operational tradeoff between false positive customer verification friction and false negative fraud losses.