End-to-end credit card fraud detection on highly imbalanced data using XGBoost + threshold tuning.

📌 Project Overview

This project focuses on detecting fraudulent credit card transactions using machine learning techniques.
The primary objective is to identify fraud cases in a highly imbalanced dataset, while minimizing false negatives — a critical requirement in real-world financial systems.

---

📊 Dataset

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

• Description:

  • Transactions made by European cardholders in September 2013
  • Highly imbalanced dataset
  • Features V1–V28 are PCA-transformed for confidentiality
  • Target variable: Class
    • 0 → Non-Fraud
    • 1 → Fraud

---

Approach & Methodology

• Performed Exploratory Data Analysis (EDA) to understand:
  • Class imbalance
  • Feature distributions
• Applied feature scaling to Time and Amount
• Used stratified train–test split to preserve class distribution
• Focused on Recall and ROC-AUC as primary evaluation metrics
• Performed threshold tuning to balance recall and precision for fraud detection

---

Models Used

• Logistic Regression (baseline)
• Random Forest
• XGBoost (primary model)

---

📈 Evaluation Metrics

Due to the imbalanced nature of the dataset, the following metrics were prioritized:

• ROC-AUC Score
• Recall (Fraud Class)
• Precision
• F1-Score
• Confusion Matrix

This ensures the model effectively identifies fraudulent transactions while controlling false positives.