SHAP-Based Feature Selection Method for Uplift Modeling

Abstract

Uplift modeling is a machine learning technique for estimating treatment effects in causal inference problems, where feature selection is crucial for reducing overfitting and computational complexity. Conventional feature selection methods often fail to meet the unique demands of uplift models. To address this, we propose a SHAP-based feature selection method that identifies important features in both treatment and control groups. Evaluated using X-learners with XGBoost as base models, our approach was tested on synthetic and real datasets against benchmark methods. Results showed that the SHAP-based method consistently outperformed existing techniques, particularly in complex real-world scenarios. It demonstrated robust performance across various feature subset sizes and excelled in handling larger feature sets and intricate data structures. These findings highlight the potential of SHAP values for accurately capturing feature importance in uplift modeling, enabling more reliable treatment effect estimation.