💳 Fraud Detection System using Machine Learning

Project Overview

This project implements an end-to-end machine learning pipeline to detect fraudulent financial transactions.
It is designed as a case-study project focusing on data understanding, feature engineering, class imbalance handling, and model evaluation.

The dataset contains over 6.3 million transactions, making this a realistic, industry-scale fraud detection problem.

---

Business Objective

• Detect fraudulent transactions with high recall
• Handle extreme class imbalance correctly
• Build a model aligned with real-world fraud behavior
• Provide actionable insights for fraud prevention

---

Dataset Description

• Total Records: 6,362,620
• Total Features: 10
• Target Variable: isFraud

Data Files

• Fraud.csv – Transaction-level dataset
• Data Dictionary.txt – Description of each column

---

Data Cleaning & Preparation

• Verified that the dataset contains no missing values
• Removed identifier columns (nameOrig, nameDest)
• Handled skewed transaction amounts using log transformation
• Reduced multicollinearity through feature engineering
• Prevented data leakage by excluding isFlaggedFraud

---

Exploratory Data Analysis (EDA)

Key findings:

• Dataset is highly imbalanced (~0.13% fraud)
• Fraud occurs mainly in TRANSFER and CASH_OUT
• Fraudulent transactions exhibit abnormal balance behavior
• Transaction amounts are heavily right-skewed

---

Feature Engineering

The following behavior-based features were created:

• Sender balance difference
• Receiver balance difference
• Zero-balance sender indicator
• Large transaction indicator
• Log-transformed transaction amount
• One-hot encoded transaction types

---

Handling Class Imbalance

• Used stratified train–test split
• Applied SMOTE only on training data
• Ensured no data leakage into the test set

---

Model Development

Model Used

• XGBoost Classifier

Why XGBoost?

• Handles non-linear fraud patterns
• Performs well on imbalanced data
• Widely used in financial fraud detection

---

Model Evaluation

Model performance was evaluated using:

• Recall (Fraud class)
• Precision and F1-score
• ROC-AUC score
• ROC curve visualization

Accuracy was not used as a primary metric due to severe class imbalance.

---

Key Fraud Drivers

• Transaction type (TRANSFER and CASH_OUT)
• Sudden depletion of sender balance
• Inconsistent receiver balance changes
• Large transaction amounts
• Transactions from zero-balance accounts

---

Fraud Prevention Recommendations

• Real-time transaction monitoring
• Step-up authentication for high-risk transactions
• Adaptive transaction limits
• Periodic model retraining and threshold tuning

---

Project Structure

1️⃣ data/

• Stores all datasets used in the project
• Raw data is never modified to maintain data integrity

Subfolders:

• raw/
  • Original dataset (Fraud.csv)
  • Data dictionary file
• processed/
  • Cleaned and feature-engineered datasets (optional / future use)

---

2️⃣ notebooks/

• Contains all Jupyter notebooks for analysis and modeling
• Notebooks are numbered to reflect the logical execution order

Notebooks:

• 01_data_loading_and_overview.ipynb

  • Data loading and understanding
  • Review of data dictionary and business context

• 02_data_cleaning_eda.ipynb

  • Data cleaning
  • Exploratory Data Analysis (EDA)
  • Class imbalance analysis

• 03_feature_engineering.ipynb

  • Creation of behavior-based fraud features
  • Encoding categorical variables
  • Reducing multicollinearity

• 04_data_splitting_and_imbalance_handling.ipynb

  • Feature–target separation
  • Stratified train–test split
  • Handling class imbalance using SMOTE

• 05_model_training.ipynb

  • Model training using XGBoost
  • Model evaluation using recall, F1-score, and ROC-AUC

---

3️⃣ models/

• Stores trained machine learning models
• Enables reuse or deployment without retraining

---

4️⃣ reports/

• Stores outputs generated from the analysis

Subfolders:

• figures/
  • Performance visualizations (ROC curve, evaluation plots)

---

5️⃣ README.md

• Provides complete project documentation
• Explains the problem, approach, results, and insights

---

6️⃣ requirements.txt

• Lists all Python dependencies required to run the project
• Ensures reproducibility of the environment

---

Conclusion

This project demonstrates a real-world, interview-ready approach to fraud detection by combining:

• Strong data understanding
• Behavior-driven feature engineering
• Proper handling of class imbalance
• Industry-standard machine learning models

---