Credit Card Fraud Detection

Project Goal: Develop a robust and financially optimized machine learning model to accurately detect fraudulent transactions within a highly imbalanced dataset, minimizing chargeback losses and false positives.

---

Key Results & Best Model Performance

The Random Forest Classifier was selected as the optimal model because it offered the best balance between identifying fraud (Recall) and minimizing false alarms (Precision), resulting in the largest net financial gain.

| Model | AUC | Precision (Class 1) | Recall (Class 1) | Financial Gain (Test Set Example) |
| Random Forest (Optimal) | 0.981 | 0.75 | 0.83 | + $56,700 |
| Logistic Regression | 0.973 | 0.06 | 0.89 | - $180,700 |
| Neural Network | 0.965 | 0.72 | 0.79 |
| K-Neighbors | 0.935 | 0.33 | 0.87 |

---

Technical Workflow

This project executes an end-to-end data science pipeline, focusing on solutions for handling imbalanced data:

1. Data Analysis and Pre-processing

• EDA: Initial analysis confirmed extreme class imbalance (less than 0.2% fraud cases) and identified high skew in the Amount feature.
• Feature Engineering: Applied a log-transform (np.log1p) to the transaction Amount to normalize its distribution, which also reduced its Variance Inflation Factor (VIF) from 13.17 to 5.03, mitigating multicollinearity.
• Scaling: Used StandardScaler on Time and log_amount to standardize inputs for distance-based and linear models.
• Class Balancing: Employed the SMOTE (Synthetic Minority Over-sampling Technique) algorithm on the training data to generate synthetic samples and address the class imbalance problem.

2. Modeling & Evaluation

• Comparative Modeling: Benchmarked performance across multiple algorithms: Logistic Regression, Decision Tree, Random Forest, Neural Network, and K-Neighbors Classifier.
• Metrics: Focused on AUC, Precision, and Recall to evaluate performance, understanding that Recall (catching fraud) and Precision (avoiding false alarms) are critical for financial applications.
• Interpretability: Used the Random Forest feature importance scores to identify the top three drivers of fraud detection: V14, V10, and V11.

3. Business Impact

• Cost-Benefit Analysis: Developed a simplified cost matrix to translate model performance (TP, FP, FN) into a tangible Net Financial Gain, justifying the Random Forest model as the financially optimal choice despite marginally lower raw AUC compared to some others.

---

Key Visualizations

---

Repository Contents

• CCFraudDetection.ipynb: The primary project notebook containing all code, analysis, and model results.
• README.md: This summary document.
• .gitignore: Specifies files (like the raw data, .csv) to be ignored by Git.
• LICENSE: MIT License governing the project's use.