💳 Credit Card Fraud Detection using Random Forest + SMOTE

---

📌 Overview

🚀 This project detects fraudulent credit card transactions using a Random Forest Classifier enhanced with SMOTE (Synthetic Minority Over-sampling Technique).
📊 Built on the Kaggle dataset, it includes preprocessing, resampling, model training, threshold tuning, evaluation, and feature importance analysis.

---

📁 Dataset Description

📦 Records: 284,807 transactions
🧬 Features:

• V1–V28: PCA-anonymized features
• Amount: Transaction amount (scaled)
• Time: Seconds since first transaction (dropped)
• Class: Target (0 = Legit ✅, 1 = Fraud ❌)

⚠️ Class Imbalance:

• Legit: 284,315 🟢
• Fraud: 492 🔴

---

📊 Exploratory Data Analysis (EDA)

📉 Visualized:

• Class distribution
• Amount distribution by class
• Correlation heatmaps (features vs Class)
• Boxplots (e.g., V14 vs Class)
• Hourly frequency of transactions
• 2D PCA scatter plot

---

⚙️ Data Preprocessing & SMOTE

🔄 Preprocessing

from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import train_test_split

df['Amount'] = StandardScaler().fit_transform(df[['Amount']])
X = df.drop(['Time', 'Class'], axis=1)
y = df['Class']

X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.3, stratify=y, random_state=42
)

🧪 Apply SMOTE

from imblearn.over_sampling import SMOTE

sm = SMOTE(random_state=42)
X_train_resampled, y_train_resampled = sm.fit_resample(X_train, y_train)

---

🌲 Model: Random Forest Classifier

from sklearn.ensemble import RandomForestClassifier

rfc = RandomForestClassifier(n_estimators=100, random_state=42, class_weight='balanced')
rfc.fit(X_train_resampled, y_train_resampled)

🧠 Trained with class_weight='balanced' to handle imbalance
🌐 100 decision trees used

---

🎯 Threshold Tuning for Better Recall

from sklearn.metrics import precision_recall_curve

y_proba = rfc.predict_proba(X_test)[:, 1]
precision, recall, thresholds = precision_recall_curve(y_test, y_proba)

optimal_threshold = 0.3
y_pred_adjusted = (y_proba > optimal_threshold).astype(int)

🎚️ Adjusted threshold from default (0.5) to improve fraud recall

---

🧪 Evaluation Metrics

🔍 Metric	📈 Value
Accuracy	99.75%
ROC AUC Score	0.99
Precision (Fraud)	~0.87
Recall (Fraud)	~0.78
F1-Score (Fraud)	~0.82

🧾 Confusion Matrix:

[[85268    27]
 [   32   116]]

📋 Classification Report:

• Class 0: Precision = 1.00, Recall = 1.00
• Class 1: Precision = 0.87, Recall = 0.78

---

🌟 Feature Importance

📌 Most important features by Random Forest:

• V17, V14, V12, V10

📊 Visualized with horizontal bar plot

---

📂 Project Structure

creditcard-fraud-detection/
│
├── notebook.ipynb              # 🔍 Full implementation and analysis
├── README.md                   # 📘 Project overview
├── images/                     # 🖼️ Visuals and plots
└── requirements.txt            # 📦 Python dependencies

---

✅ Key Takeaways

✔️ Random Forest + SMOTE = Powerful combo for imbalanced fraud detection
📈 Threshold tuning improves recall for fraud cases
📊 Features V14, V17, V12, and V10 are highly informative
💡 Easy to interpret, scalable, and reproducible

---

🛠️ Future Improvements

📌 Try alternative models:

• XGBoost 🌲
• LightGBM ⚡
• Logistic Regression 📈

🧪 Add:

• GridSearchCV for hyperparameter tuning
• Real-time deployment using Flask / Gradio / Streamlit

---

📦 Dependencies

numpy
pandas
matplotlib
seaborn
scikit-learn
imblearn

---

📜 License

MIT License © 2025 Anton Atef

---

🤝 Contributions

👨‍💻 Feel free to fork, clone, and submit pull requests!
📬 Suggestions and issues are welcome anytime!

---

📬 Contact

📧 Email: tony.atef.954@gmail.com

---