// the find
Fraud-Detection-Handbook/fraud-detection-handbook
Reproducible Machine Learning for Credit Card Fraud Detection - Practical Handbook
A free, work-in-progress online book on machine learning for credit card fraud detection, written by ULB's Machine Learning Group and Worldline researchers. It's aimed at students and practitioners who want a practical, code-first treatment of fraud detection rather than a survey paper — every chapter after the background section is a runnable Jupyter notebook.
The notebooks are genuinely reproducible: they include a synthetic transaction simulator so you can run the full pipeline without needing a real (and legally restricted) card dataset. It covers the parts most ML tutorials skip for fraud specifically — prequential/time-aware validation instead of random k-fold, top-k and threshold-free metrics instead of plain accuracy, and cost-sensitive learning tied to actual fraud amounts. The authors have a decade of applied research behind this (Worldline is a real payment processor), so the imbalanced-learning and sequential-modeling chapters reflect production concerns, not just textbook recipes.
Chapter 8 (interpretability) was never published and the repo has had no commits since early 2024, so treat this as a frozen draft, not a maintained project. It's a Jupyter Book, not a library — there's no installable package or API, so you're copying notebook code rather than importing anything. The pinned toolchain (Sphinx 4.2, jupyter-book 0.11.2) is old enough that building the book locally will likely hit dependency conflicts on a current Python environment. Real-world data chapters use a private Worldline dataset you don't have access to, so those notebooks aren't actually reproducible for an outside reader, only the simulated-data ones are.