Imagine building a fraud detection model that achieves 99.83% accuracy but fails to detect a single fraudulent transaction.
That's the problem I wanted to explore with FraudLens ML.
The project examines credit-card fraud detection using a benchmark dataset containing 284,807 transactions, including only 492 fraudulent transactions.
What I explored
- Logistic Regression, Random Forest, and XGBoost
- Class imbalance and random oversampling
- Precision, recall, F1, PR-AUC, and MCC
- Leakage-safe train/test splitting
- Interactive Streamlit visualisation
Why it matters
Accuracy alone can hide poor minority-class performance. A fraud detector must be evaluated on how many fraud cases it catches and how many legitimate transactions it incorrectly flags.
The project also distinguishes published research results from independently generated experimental results.
Limitations
The dashboard presents research benchmarks and supports uploading experimental results. Published benchmark scores should not be mistaken for results reproduced by my implementation.
GitHub repository: https://github.com/tasnem-tech/fraudlens-ml
What evaluation metric do you prioritise for highly imbalanced classification problems?
Top comments (0)