FLEX-TALENT: AN EXPLAINABLE AND FAIR MACHINE LEARNING FRAMEWORK FOR HIGH-STAKES LEADERSHIP-PIPELINE EVALUATION
World Journal of Management Science · 2024
DOI: 10.61784/wms4121 ↗Kurz gefasst
Fachartikel in der Zeitschrift World Journal of Management Science, 2024. Verfasst von ZiMeng Wang und 1 weiteren Autor:innen. Die Arbeit wurde bisher 2-mal zitiert (Crossref). Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Machine learning can prioritize candidates for leadership pipelines, but historical promotion records encode organizational policy rather than objective executive potential. This distinction is critical in high-stakes talent evaluation, where a technically accurate model may reproduce label bias, obscure unequal error burdens, or invite automation bias. This paper presents FLEX-Talent, a construct-aware, budget-constrained framework that combines feature governance, training-only fairness reweighting, predeclared validation gates, group calibration audits, intersectional testing, and explanation-stability checks. The framework is evaluated on a real public HR promotion dataset containing 54,808 employees and 4,668 historical promotions. Gender is never used at inference; it is retained only for auditing and reweighting. Five candidate models are compared under a fixed 10% shortlist budget. The validation rule selected a reweighted CatBoost model, which achieved 0.9134 ROC-AUC and 0.6116 average precision on an untouched test set. However, its absolute gender equal-opportunity difference increased from 0.0138 on validation to 0.0578 on test, slightly exceeding the predeclared 0.05 gate; a 95% bootstrap interval of [0.0060, 0.1497] reveals substantial fairness uncertainty. Rather than treating this as a hidden failure, FLEX-Talent makes fairness generalization a first-class release decision. Global and local SHAP evidence is accompanied by bootstrap stability analysis and explicit non-causal warnings. The resulting contribution is not an automated promotion system, but a reproducible governance architecture for human review, appeal, and post-deployment monitoring.
Zitieren
APA
Wang, Z., & Shen, Z. (2024). FLEX-TALENT: AN EXPLAINABLE AND FAIR MACHINE LEARNING FRAMEWORK FOR HIGH-STAKES LEADERSHIP-PIPELINE EVALUATION. World Journal of Management Science. https://doi.org/10.61784/wms4121
Vancouver
Wang Z, Shen Z. FLEX-TALENT: AN EXPLAINABLE AND FAIR MACHINE LEARNING FRAMEWORK FOR HIGH-STAKES LEADERSHIP-PIPELINE EVALUATION. World Journal of Management Science. 2024. doi:10.61784/wms4121
BibTeX
@article{Wang2024,
author = {Wang, ZiMeng and Shen, ZiJian},
title = {FLEX-TALENT: AN EXPLAINABLE AND FAIR MACHINE LEARNING FRAMEWORK FOR HIGH-STAKES LEADERSHIP-PIPELINE EVALUATION},
journal = {World Journal of Management Science},
year = {2024},
doi = {10.61784/wms4121}
}
EUVIMED