mdsajjadullah/fairxai-shap-eds-fraud-detection

--- tags: - fraud-detection - tabular-classification - xgboost - shap - explainable-ai - fairness - bias-detection - model-interpretability --- # 🧠 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--- 标签： - 欺诈检测 - 表格分类 - XGBoost - SHAP - 可解释AI（Explainable AI） - 公平性 - 偏差检测 - 模型可解释性 --- # 🧠 FairXAI-欺诈检测（FairXAI-FraudDetection） ## 🚀 项目概述 本方案为一款面向信用卡欺诈检测的端到端可解释AI（Explainable AI）流水线，不仅优化预测性能，还可同步评估模型决策与解释结果的公平性。 ## ⚠️ 问题痛点 传统欺诈检测系统多为黑盒模型，仅在预测层面（准确率、F1值）评估公平性，无法衡量不同客户群体间的解释结果是否保持一致。 ## 💡 核心贡献 **解释差异评分（Explanation Disparity Score，EDS）** 一种全新的量化指标，用于评估基于SHAP的解释结果在不同客户消费画像间是否保持一致。 ## 📊 实验结果 - XGBoost：AUC-ROC 0.9802 | AUC-PR 0.8743 | F1 0.8195 - 中等消费群体：召回率最低（0.667） - 高端消费群体：综合性能最优（F1值0.894） - 高消费群体：呈现独特的解释模式（V4与V14特征） - EDS-F1：0.1792 → 预测差异度 - EDS-SHAP：0.0345 → 解释一致性 ## 🔍 核心发现 该模型在全局层面的欺诈解释逻辑保持一致，但在不同客户群体间的欺诈检测性能存在显著差异。 ## 🛠️ 实现方法 - 所用模型：XGBoost、随机森林、逻辑回归 - 可解释性方法：SHAP - 公平性评估：基于5类消费画像的EDS指标 ## 📁 数据集信息 信用卡欺诈检测（Credit Card Fraud Detection，ULB） 包含284,807笔交易记录、492个欺诈案例，共30个特征 ## 🔗 相关资源 - Kaggle 数据集链接：https://www.kaggle.com/datasets/mdsajjadullah/fairxai-fraud-detection-and-fairness-audit - GitHub 仓库链接：https://github.com/MdSajjadUllah/FairXAI-FraudDetection