Ensemble SHAP-Based Feature Selection for Credit Card Fraud Detection

  
Author Huanjing Wang

 
Co-Author(s) Qianxin Liang; Taghi M. Khoshgoftaar

 
Abstract Feature selection plays a critical role in enhancing the performance and interpretability of machine learning models, particularly in high-dimensional datasets. In this study, we introduce an ensemble SHAP-based feature selection approach that integrates SHAP value rankings produced by three distinct base learners: Decision Tree, Random Forest, and XGBoost. The ensemble method combines individual SHAP value rankings using a median approach, resulting in a more stable and reliable feature selection strategy. We evaluate the effectiveness of the proposed approach by training Random Forest classifier on various subsets of top-ranked features, as well as on the full feature set. Model performance is assessed using the Area Under the Precision-Recall Curve (AUPRC). The experiments in this study employ the Kaggle Credit Card Fraud Detection Dataset. The results demonstrate that using the top 15 features, selected by the ensemble method, achieves performance nearly equivalent to that of the full feature set, significantly reducing dimensionality with minimal loss in predictive accuracy. These findings suggest that the ensemble SHAP-based feature selection technique effectively balances model accuracy and feature efficiency. This method holds promise for applications requiring interpretable and computationally efficient models.

 
Keywords Ensemble, SHAP, Feature Selection, Credit Card Fraud Detection, Machine Learning
   
    Article #:  RQD2025-162
 

Proceedings of 30th ISSAT International Conference on Reliability & Quality in Design
August 6-8, 2025