Credit Card Fraud Detection: Practical Exploration of Hybrid Machine Learning and Deep Learning Models

This project addresses the extreme class imbalance problem in credit card fraud detection by building an end-to-end system. It integrates multiple algorithms including logistic regression, random forest, XGBoost, feedforward neural networks, and autoencoders. Using techniques like SMOTE oversampling and dynamic weighted ensemble learning, it maintains high recall while controlling false positive rates, providing a complete technical framework for financial fraud detection.

Global annual credit card fraud losses amount to tens of billions of US dollars. The core challenge is extreme data imbalance (fraudulent transactions usually account for less than 0.1%), causing traditional models to tend to favor normal transactions. The project uses a European cardholder credit card transaction dataset, which contains 30 features (V1-V28 are PCA anonymized features, Time, Amount, and Class are original features), with fraudulent samples accounting for approximately 0.17%.

1. Feature Standardization: Scale Time and Amount using StandardScaler to mean 0 and variance 1; 2. Stratified Sampling Split: 80% training set + 20% test set, maintaining consistent fraud ratio; 3. SMOTE Oversampling: Generate synthetic samples via interpolation between minority class samples to alleviate class imbalance.

• Traditional ML Models: Logistic Regression (dynamic threshold optimization), Random Forest (class weight adjustment + feature importance analysis), XGBoost (scale_pos_weight for imbalance handling + regularization); - Deep Learning Models: Feedforward Neural Network (64/32/16 hidden layers + Dropout + early stopping), Autoencoder (unsupervised learning of normal transaction patterns, identifying fraud via reconstruction error).

Dynamically assign weights based on PR-AUC, integrating prediction results from logistic regression, random forest, XGBoost, and neural networks. The formula is: Ensemble Probability = w₁×LR + w₂×RF + w₃×XGB + w₄×NN. Advantages: Reduce bias of single models, improve generalization ability, and flexibly balance precision and recall.

Evaluation metrics include precision, recall, F1-score, ROC-AUC, PR-AUC, and confusion matrix; Visualization content: class distribution chart, confusion matrix heatmap, ROC/PR curve comparison, feature importance bar chart, neural network training curve, etc., to intuitively display model performance.

The project provides a complete technical framework for financial fraud detection. Its methodology can be transferred to scenarios such as insurance fraud, money laundering identification, and account theft detection; it also has reference significance for fields like medical rare disease detection, industrial defect detection, and cybersecurity intrusion detection.