Fraud Detection ML Pipeline

Project Overview: Implementation of a machine learning pipeline for fraud detection using the Credit Card Fraud Detection dataset.
The pipeline is inspired by arxiv:2208.11900 in the sense that I examined the models they used to confirm that the dataset is predictable. This project focuses on building a reproducible, fully end-to-end ML pipeline without sampling or altering the dataset.

Project Goals

• Validate and preprocess the full dataset
• Handle missing values robustly
• Apply scaling with consideration for skewed features
• Build and train multiple ML models for fraud detection
• Track experiments with MLflow for full reproducibility

Dataset Assumptions

• PCA-transformed features: Original features have been transformed using Principal Component Analysis (V1-V28). Feature names are not semantically meaningful.
• No feature engineering: For security reasons, features are ambigious, thus only scaling is performed. No additional feature creation or selection.
• Null values: Handled defensively (median imputation) though none are expected.
• Scaling: StandardScaler or RobustScaler is applied depending on feature skew.

Implementation Details

Preprocessing Pipeline

1. Validation: Schema checks, null checks, shape consistency, target value validation.
2. Imputation: Median for numeric nulls.
3. Scaling: StandardScaler or RobustScaler depending on skew.
4. No feature engineering: Dimensions preserved throughout.

Model Pipeline

• Multiple classical ML models are used to confirm predictability.
• Model configurations are informed by the original paper but not copied word-for-word.
• Hyperparameters are configurable via config.yaml.

Project Structure

FRAUD_DETECTION/
├── config.yaml           # Configuration file
├── main.py               # Main entry point
├── requirements.txt      # Python dependencies
├── README.md             
├── data/
│   ├── raw/              # Raw data (provided externally)
│   ├── interim/          # Intermediate validated data
│   └── processed/        # Final processed data
├── src/
│   ├── ingest.py         # Training loop + MLflow logging
│   ├── preprocessor.py   # Orchestrates validation + transformation
│   ├── validator.py      # Data validation class
│   ├── transformer.py    # Scaling (Standard/Robust)
│   ├── model.py          # ML models
│   ├── evaluate.py       # Evaluation metrics
│   ├── training_utils.py # Training utility functions
│   └── train.py          # Training pipeline + MLflow logging
├── artifacts/            # Saved models, transformers (Such as scalers...)
├── logs/                 # Log files
└── tests/                # Unit tests

Reproducibility Instructions

• Python 3.8+
• Install dependencies: pip install -r requirements.txt
• Run training: python main.py --mode train
• Track experiments: mlflow ui → http://localhost:5000

All randomness is controlled via seeds in config.yaml, and fitted scalers/models are saved for reproducibility.

Key Notes

• This project does not claim full replication of the original paper.
• The paper was consulted to estimate which models performed best on this dataset WITHOUT sampling.
• No sampling was applied — the entire dataset is used for training and evaluation.

Code Quality

• Type hints and docstrings throughout
• Comprehensive error handling
• Unit tests included
• Logging and MLflow tracking for all experiments