Machine Learning · 2026
FraudRadar Detection System
A machine learning system that flags suspicious credit card transactions and assigns risk scores. It helps fraud teams prioritize which transactions to review. The Extra Trees model reached 94.83% precision and 77.46% recall.
- 94.83%PrecisionExtra Trees model on test data
- 77.46%RecallFraud cases successfully detected
- 55/58True fraud alertsThree alerts were false positives
Case study contents
Objective
Detect fraudulent transactions from highly imbalanced data. The model turns predictions into simple risk scores to support human investigation.
Results and limitations
The model reached 94.83% precision and 77.46% recall on the test set. Out of 58 alerts, 55 were true fraud and 3 were false. Main limitation, the dataset covers only two days of transactions and most features are anonymous.
Visual evidence
Technical details
Open implementation details
Role and contribution
Worked on the entire project alone. Cleaned data, compared several models, tuned the decision threshold, and built a Streamlit dashboard. Also wrote unit tests and a CI pipeline.
Methodology
Cleaned the data and removed duplicates. Compared several imbalance handling techniques and models. Selected the best model and tuned the decision threshold. Added calibration checks and drift monitoring.
Technologies
- Python
- scikit-learn
- XGBoost
- LightGBM
- CatBoost
- SMOTE
- Streamlit
- Pandas
- pytest
- GitHub Actions