XAI-Driven Financial Fraud Detection System

An end-to-end Machine Learning framework for financial fraud detection that integrates
anomaly detection, ensemble learning, and Explainable AI (XAI) to achieve high
accuracy while maintaining model transparency and auditability.

This project is based on a peer-reviewed research publication and focuses on building
trustworthy, interpretable ML systems suitable for real-world financial environments.

🔍 Problem Statement

Financial fraud detection systems face three major challenges:

• Severe class imbalance (fraud cases are rare)
• Constantly evolving fraud patterns
• Lack of explainability, making ML models difficult to trust and regulate

This project addresses these challenges by combining:

• SMOTE-based data balancing
• Anomaly detection models
• Ensemble classifiers
• Explainable AI techniques (SHAP, LIME)

🧠 System Architecture

Pipeline Overview

1. Data Preprocessing

   • Data cleaning and feature engineering
   • Class imbalance handling using SMOTE

2. Anomaly Detection Layer

   • Autoencoder-based reconstruction error
   • Isolation Forest for outlier detection

3. Classification Layer

   • Random Forest
   • XGBoost (primary high-performing model)

4. Explainability Layer

   • SHAP for global and local feature importance
   • LIME for instance-level explanations

This layered architecture ensures accuracy, robustness, and interpretability.

📊 Key Results (Research-Backed)

• SMOTE improved fraud recall from 0.44 → 0.76
• XGBoost delivered the best balance between precision and recall
• SHAP enabled transparent feature-level decision tracing

📁 Repository Structure

XAI-Financial-Fraud-Detection/
│
├── src/ # Core ML pipeline
│ ├── preprocessing.py
│ ├── anomaly.py
│ ├── classifier.py
│ ├── explainability.py
│ └── config.py
│
├── data/ # Transaction datasets
├── models/ # Trained models & encoders
├── outputs/ # Evaluation results & plots
│
├── publications/ # Research paper & poster
├── requirements.txt
└── README.md

🚀 How to Run

Install dependencies:

pip install -r requirements.txt

Run the main pipeline:
python src/main.py

Launch the dashboard (if applicable):
python fraud_dashboard.py

📄 Research Publication

This implementation is based on the peer-reviewed paper:

“A Machine Learning Framework for Financial Fraud Detection Using Explainable Artificial Intelligence Techniques”
Published in International Journal of Computer Sciences and Engineering (IJCSE), 2025

🔗 DOI: https://doi.org/10.26438/ijcse/v13i5.1725

👥 Authors & Contributions

Ashish Seru – ML pipeline design, model development, evaluation

Archit Mehrotra – Co-developer, experimentation & validation

Tanisha Gotadke – Documentation & research synthesiss

Faculty mentors – Research supervision

This repository represents the engineering implementation of the published research.

🛠 Tech Stack

Python

Scikit-learn

XGBoost, Random Forest

Isolation Forest

SHAP, LIME

Pandas, NumPy

Streamlit

🎯 Why This Project Matters

This project goes beyond a toy ML implementation and reflects real-world financial system requirements:

• Production-oriented ML design with modular, maintainable pipeline components
• Explainable AI (XAI) to support transparency and regulatory compliance
• Research-backed methodology, validated through a peer-reviewed publication
• Clear separation of ML stages (preprocessing, modeling, evaluation, explainability)
• Industry relevance, aligned with fraud detection use cases in banking and fintech systems

📬 Contact

Ashish Seru
GitHub: https://github.com/AshishSeru

LinkedIn: https://www.linkedin.com/in/ashishseru/