mdsajjadullah/fairxai-shap-eds-fraud-detection

🧠 FairXAI-FraudDetection

🚀 Overview

An end-to-end Explainable AI pipeline for credit card fraud detection that goes beyond prediction performance to evaluate fairness in both model decisions and explanations.

⚠️ Problem

Traditional fraud detection systems act as black boxes and measure fairness only at the prediction level (accuracy, F1).
They fail to assess whether explanations are consistent across different customer groups.

💡 Contribution

Explanation Disparity Score (EDS)
A novel metric to measure whether SHAP-based explanations remain consistent across customer spending profiles.

📊 Results

• XGBoost: AUC-ROC 0.9802 | AUC-PR 0.8743 | F1 0.8195
• Moderate Spenders: lowest recall (0.667)
• Premium Spenders: highest performance (F1 0.894)
• High Spenders: distinct explanation pattern (V4 vs V14)
• EDS-F1: 0.1792 → prediction disparity
• EDS-SHAP: 0.0345 → explanation consistency

🔍 Key Insight

The model is consistent in how it explains fraud globally, but inconsistent in how well it detects fraud across customer groups.

🛠️ Methods

• Models: XGBoost, Random Forest, Logistic Regression
• Explainability: SHAP
• Fairness: EDS across 5 spending profiles

📁 Dataset

Credit Card Fraud Detection (ULB)
284,807 transactions · 492 fraud cases · 30 features

🔗 Resources

• Kaggle : https://www.kaggle.com/datasets/mdsajjadullah/fairxai-fraud-detection-and-fairness-audit
• GitHub: https://github.com/MdSajjadUllah/FairXAI-FraudDetection