Practical Guide to Credit Card Fraud Detection: A Complete Machine Learning Pipeline from Data Exploration to Multi-Model Comparison

This article introduces the open-source project fraud-detection-ml, which addresses the problem of extreme class imbalance in credit card fraud detection by building a complete machine learning pipeline from data exploration to model deployment. The project covers feature engineering, application of SMOTE oversampling technology, and comparison of multiple models (logistic regression, random forests, XGBoost, neural networks), providing practical references for financial risk control.

Credit card fraud causes tens of billions of dollars in losses globally each year. Detection faces extreme class imbalance (fraudulent transactions account for <0.1%), rendering the accuracy metric ineffective. The project uses the Creditcard dataset of two days of transactions from European cardholders, which includes PCA-anonymized features V1-V28, Amount, and Time. Data characteristics: Fraudulent transactions have concentrated amount distribution and time clustering; the class distribution is extremely imbalanced (fraud accounts for 0.17%), so metrics like precision and recall need to be focused on.

Feature Engineering: 1. Log transformation of amount to compress long-tail distribution; 2. Extract hour from time and perform periodic encoding (sine/cosine); 3. Scale amount features with RobustScaler (robust to outliers). Class Imbalance Handling: Apply SMOTE only on the training set to generate synthetic minority samples (avoid data leakage). Model Selection: Baseline logistic regression (interpretable), random forests (nonlinear interactions + feature importance), XGBoost (tuning + SHAP interpretation), MLP (nonlinear mapping). Tuning: RandomizedSearchCV + stratified K-fold cross-validation (maintain class ratio).

Evaluation Metrics: Confusion matrix (focus on missed fraud FN and false positives FP), classification report (precision/recall/F1), ROC-AUC (overall discrimination ability), PR-AUC (more sensitive to imbalanced scenarios). Threshold Tuning: Select based on business needs (low threshold for high recall, high threshold for high precision). Feature Importance: Random forest feature ranking, SHAP value analysis for XGBoost, revealing key feature contributions.

1. Modular design: Separate data loading, exploration, preprocessing, training, and evaluation; 2. Centralized configuration management: Unified parameters in config.py; 3. Output management: Automatically save EDA charts, model comparison graphs, etc., to the outputs directory; 4. Colab support: Provide cloud notebooks to lower the entry barrier.

Insights: 1. Class imbalance requires combining technology (SMOTE) and business (threshold selection); 2. Model selection serves business objectives; 3. Evaluation metrics align with business costs; 4. Interpretability is essential (e.g., SHAP values). Limitations: Did not consider transaction temporal characteristics (e.g., historical behavior, correlation of multiple transactions in a short time); the dataset is PCA-anonymized, missing contextual information like merchant type and geographic location. Improvements: Introduce temporal features and supplement real business context data.

The fraud-detection-ml project provides a complete credit card fraud detection pipeline, covering key links such as data exploration, feature engineering, imbalance handling, multi-model comparison, and evaluation. It is a valuable learning resource for beginners in machine learning for financial risk control and practitioners in imbalanced scenarios.