Credit Card Fraud Detection
Classification
Machine Learning
Classifying real fraudulent transactions under extreme class imbalance, on the classic ULB benchmark.
Problem
Identify fraudulent credit card transactions from European cardholders (September 2013), where fraud is a tiny fraction of all transactions.
Approach
Handling extreme class imbalance and choosing metrics suited to the problem, comparing classification models.
Technologies
Python scikit-learn Imbalanced classification XGBoost Optuna SHAP Streamlit
Results
With a 578:1 imbalance across 284,807 transactions, the model reached PR-AUC 0.826, recall 0.811 and precision 0.951 at a 0.264 threshold, chosen from the precision-recall curve rather than the 0.5 default. PR-AUC was used instead of ROC-AUC, which is misleading under extreme imbalance.