https://github.com/gss-18/fraud-detection

🚀 Credit Card Fraud Detection using Machine Learning & XGBoost. Evaluating models on an imbalanced dataset, using SMOTE, ROC-AUC analysis, and finding the best fraud detection approach. 📊🔍
https://github.com/gss-18/fraud-detection

classification credit-card-fraud data-science fraud-detection imbalanced-data machine-learning python sklearn smote xgboost

Last synced: about 2 months ago
JSON representation

🚀 Credit Card Fraud Detection using Machine Learning & XGBoost. Evaluating models on an imbalanced dataset, using SMOTE, ROC-AUC analysis, and finding the best fraud detection approach. 📊🔍

• Host: GitHub
• URL: https://github.com/gss-18/fraud-detection
• Owner: gss-18
• Created: 2025-03-13T18:06:23.000Z (about 2 months ago)
• Default Branch: main
• Last Pushed: 2025-03-13T18:28:39.000Z (about 2 months ago)
• Last Synced: 2025-03-13T19:38:00.159Z (about 2 months ago)
• Topics: classification, credit-card-fraud, data-science, fraud-detection, imbalanced-data, machine-learning, python, sklearn, smote, xgboost
• Language: Jupyter Notebook
• Homepage:
• Size: 152 KB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# 💳 Credit Card Fraud Detection Using XGBoost  

## 🚀 **Project Overview**

This project aims to detect fraudulent credit card transactions using machine learning. We experimented with multiple models and **found that the default XGBoost model performed the best**. Despite extensive hyperparameter tuning, default XGBoost provided the best balance of **precision, recall, and F1-score** for fraud classification.

---

## 📂 **Dataset Information**

- **Dataset:** [Kaggle - Credit Card Fraud Detection](https://www.kaggle.com/datasets/mlg-ulb/creditcardfraud)  

- **Total Transactions:** 284,807  

- **Fraud Cases:** 492 (~0.17% of total transactions)  

- **Non-Fraud Cases:** 284,315  

- **Features:**  

  - 30 anonymized features (`V1` to `V28`) derived using **PCA**  

  - `Time`: Seconds elapsed since the first transaction  

  - `Amount`: Transaction amount  

  - `Class`: **Target variable (1 = Fraud, 0 = Non-Fraud)**  

---

## 🔬 **Project Methodology**

1. **Data Preprocessing**

   - Standardized the `Amount` feature.

   - Used **SMOTE (Synthetic Minority Oversampling Technique)** to balance fraud cases in training data.

2. **Model Training & Testing**

   - Tried different models (**Logistic Regression, Decision Tree, LightGBM, Neural Networks**).

   - **XGBoost outperformed all others**.

   - Used **GridSearchCV** to fine-tune hyperparameters.

3. **Final Model Selection**

   - Compared **Default XGBoost vs Hyperparameter-Tuned XGBoost**.

   - Surprisingly, **Default XGBoost gave better results**.

   - **Final decision:** Use **default XGBoost** with minor threshold tuning.

---

## 📊 **Model Performance Comparison**

| **Model**             | **Precision (Fraud)** | **Recall (Fraud)** | **F1-Score (Fraud)** |

|---------------------|--------------------|-----------------|-----------------|

| **Default XGBoost** | **0.79**           | **0.85**        | **0.82**        |

| **Tuned XGBoost**   | **0.35**           | **0.88**        | **0.50**        |

✅ **Default XGBoost was superior** because:  

- It had a **higher F1-score (0.75 vs 0.50)**.  

- It maintained a **strong fraud recall (0.86)** without excessive false positives.  

- The tuned model had **very low precision (0.35)**, leading to **too many false alarms**.

---

## 🚀 **Final Model: Default XGBoost**

### **🔹 Why Did Hyperparameter Tuning Fail?**

- **`scale_pos_weight=7` hurt precision** by prioritizing fraud cases too much.  

- **`max_depth=5 & n_estimators=150` overfitted the training data**.  

- **Default settings provided a well-balanced model**, making extra tuning unnecessary.

## 📈 **Final ROC-AUC & Precision-Recall Curve**

After evaluating the model, we analyzed the **ROC-AUC Curve** and **Precision-Recall Curve** to measure the fraud detection performance.

### **🔹 ROC-AUC Score (Receiver Operating Characteristic - Area Under Curve)**

- **ROC-AUC Score:** **0.97**

- A high **ROC-AUC score indicates strong discrimination** between fraud and non-fraud cases.

- The model effectively **minimizes false positives while capturing fraudulent transactions**.

### **🔹 Precision-Recall Curve Analysis**

- **Average Precision (AP) Score:** **0.77**

- This curve is especially useful for highly **imbalanced datasets** like fraud detection.

- The model maintains **high precision at lower recall values**, ensuring fraud is detected efficiently.

| **Metric**  | **Value** |

|------------|----------|

| **ROC-AUC** | 0.99 |

| **PR AUC**  | 0.85 |

✅ **These results confirm that the model successfully detects fraud with minimal false positives.**

---

## 🎯 **Key Takeaways**

1️⃣ **Default XGBoost performed better than hyperparameter-tuned versions.**  

2️⃣ **High recall (86%) ensures most fraud cases are detected.**  

3️⃣ **Precision (67%) is reasonable, balancing fraud detection and false alarms.**  

4️⃣ **Over-tuning `scale_pos_weight` and deep trees led to overfitting.**  

5️⃣ **Sometimes, default settings work best!**  

---

## 📌 **Next Steps for Further Improvement**

🔹 **Test Ensemble Learning** (XGBoost + LightGBM + Neural Network)  

🔹 **Try a Hybrid Sampling Approach** (SMOTE + Undersampling)  

🔹 **Experiment with AutoML-based tuning for even better performance**  

---