Integrating F-filter and K-Means SMOTE to enhance LGWUM-based turnover prediction

Abstract

High employee turnover poses a significant challenge for organizations. While uplift modeling offers a prescriptive analytics approach by estimating differential treatment effects to optimize retention programs, its performance is often hindered by irrelevant features and imbalanced class distributions. To address these issues, this study proposes an employee turnover model utilizing lai’s generalized weighted uplift method (LGWUM), enhanced with F-Filter feature selection and K-Means SMOTE for a refined feature space and balanced treatment representation. Evaluated across three HR datasets with four engineered uplift classes (CN, CR, TN, TR), the integrated framework significantly improves uplift performance, yielding increased Qini coefficients of 0.0755 and 0.0870 on Datasets 2 and 3, respectively. Furthermore, top-decile probability distribution analysis confirms a clearer separation between positive and negative responders, with XGBoost demonstrating the most robust and reliable uplift discrimination across the models.