🏦 Credit Card Fraud Detection

📌 Overview

This project focuses on detecting fraudulent credit card transactions using machine learning techniques. Given the extreme imbalance in fraud transactions, methods like SMOTE (Synthetic Minority Over-Sampling Technique) and Logistic Regression have been employed to enhance model performance.

📂 Dataset Description

The dataset consists of credit card transactions over a two-day period in September 2013, containing:

• 284,807 transactions
• 492 fraud cases (0.172% of total transactions)

🔑 Features:

• V1 to V28: PCA-transformed features (due to confidentiality).
• Time: Elapsed time since the first transaction.
• Amount: Transaction amount.
• Class: Fraud status (0 = Non-fraud, 1 = Fraud).

---

🛠️ Preprocessing Techniques

⚖️ Handling Class Imbalance:

• Since fraudulent transactions are extremely rare (0.17%), we applied SMOTE to balance the dataset.
• Fraudulent instances increased from 469 to 65,598 post-balancing.

🏗️ Feature Engineering:

• Weight of Evidence (WOE): Used to transform categorical variables into continuous values.
• Information Value (IV): Selected the most predictive features (IV > 0.3).

---

📊 Exploratory Data Analysis

• 📈 Correlation heatmaps to identify relationships between features.
• 📉 Distribution analysis to understand fraud vs. non-fraud transactions.
• 🛑 Fraud vs. Non-Fraud Imbalance visualization.

---

🤖 Machine Learning Model

The primary model used is Logistic Regression, a widely accepted model for fraud detection.

🔍 Model Evaluation:

• ROC Curve 📈: Visualizing classification performance.
• Fraud Score Calculation 🎯: Transformed log-odds into a fraud risk score (0-100), categorizing transactions as:
  • 🟢 No Risk
  • 🟡 Low Risk
  • 🟠 Moderate Risk
  • 🔴 High Risk