Credit Card Transaction Classifier

A machine learning pipeline designed to identify fraudulent transactions with high precision and recall. This project focuses on handling extreme class imbalance and preventing data leakage through a robust pipeline architecture.

Project Structure

The project is organised as follows:

credit-card-fraud-ml/
│
├── data/           # Source CSV data (e.g. creditcard.csv)
├── docs/           # Technical report and documentation
├── figures/        # Exported Precision-Recall curves and EDA plots
├── models/         # Serialised champion model (.pkl)
├── notebooks/      # EDA and benchmarking experiments
├── src/            # Python scripts and utilities
├── requirements.txt
└── README.md

Methodology

• Anti-Leakage Pipeline: Utilises imblearn.pipeline.Pipeline to ensure that the RobustScaler is fitted strictly on training data, eliminating look-ahead bias.
• Cost-Sensitive Learning: Addresses the extreme 0.17% fraud imbalance by utilising scale_pos_weight in XGBoost and class_weight='balanced' in Linear models. This proved more effective than naive scaling during experimentation.
• Metric Focus: Prioritises AUPRC (Area Under the Precision-Recall Curve) to ensure high detection (Recall) while minimising false alarms (Precision).
• Model Suite: Compares Logistic Regression, Random Forest, Hist-Gradient Boosting, XGBoost, and Linear SVM.

How to Run

1. Install dependencies:

   pip install -r requirements.txt

2. Explore the Research: Review notebooks/eda.ipynb for data insights and notebooks/experiments.ipynb to see the benchmarking process and leakage analysis.

3. Train the Champion Model: Execute the training suite to benchmark all models and serialise the best performer to the models/ directory:

   python src/train.py

4. Verify Results: Run the final verification script to load the serialised pipeline and test it against the held-out dataset. This will also save a final verification plot to figures/:

   python src/test.py