Practical Guide to Credit Card Fraud Detection: In-depth Comparison of Random Forest and Class Imbalance Handling Methods

This project is a hands-on initiative for machine learning beginners, developed by WangareCeline. It deeply explores class imbalance issues in credit card fraud detection by comparing three methods: baseline random forest model, SMOTE oversampling, and class weight adjustment. Using the Kaggle Credit Card Fraud Dataset, the project reveals the important conclusion that simple baseline models may outperform complex strategies in specific scenarios, providing a reference for similar problems.

Financial fraud detection faces extreme class imbalance (fraudulent transactions usually account for less than 2%), making accuracy metrics misleading—recall and F1-score should be prioritized. This project uses Kaggle's Credit Card Fraud Detection Dataset, which contains 10,000 records: 9,849 normal transactions (98.5%) and 151 fraudulent transactions (1.5%). Features include transaction amount, time, merchant category, risk indicators, etc.

Data preprocessing steps: No missing value cleaning, label encoding for merchant_category, remove transaction_id; split into 80/20 training/test sets (random_state=42). Model strategies: 1. Baseline model: Standard random forest with 100 trees; 2. SMOTE oversampling: Generate synthetic samples via interpolation between minority class samples; 3. Class weight adjustment: Assign higher weights to the fraud class, modify loss function to penalize misclassification of minority class.

Comparison of experimental results:

Model	Precision	Recall	F1 Score
Baseline	1.00	0.61	0.76
SMOTE	0.26	0.61	0.36
Class Weight	1.00	0.55	0.71
Key findings: The baseline model has the highest F1 score; SMOTE's precision drops sharply (many false positives); class weight adjustment leads to a slight decrease in recall. The baseline model has zero false positives (0 normal transactions misclassified as fraud), making it highly valuable for practical applications.

Practical insights: 1. Complex techniques are not necessarily better than simple baselines; 2. For imbalanced datasets, prioritize recall/F1 score; 3. Strategy effectiveness depends on data characteristics (SMOTE may perform poorly on small datasets); 4. Features built with domain knowledge (e.g., location_mismatch) are more predictive. Project significance: Provides a standardized practice template for beginners and emphasizes a scientific and rigorous experimental attitude.

The project uses Python data science ecosystem tools: Python3, Pandas (data processing), NumPy (numerical computation), Scikit-learn (modeling/evaluation), Imbalanced-learn (SMOTE), Matplotlib/Seaborn (visualization), Jupyter Notebook (development environment).