Central Asian Journal of STEM, 2026
Leakage-Controlled Machine Learning Evaluation for Workforce Attrition Prediction in Automated Production Control Systems: A Methodological Standard
Turin Polytechnic University
Аннотация · Abstract
Abstract: We demonstrate that incorrect SMOTE placement before train-test splitting inflates PR-AUC by 49.2% (0.589→0.879)—a systematic methodological error prevalent in published HR analytics literature. Under a rigorous 5×3 nested cross-validation protocol, logistic regression achieves CV PR-AUC=0.624±0.058, hold-out PR-AUC=0.576 [95% CI: 0.446–0.697], recall@20%=0.574 (feed-forward gain α), Brier=0.101, MCC=0.496. A six-algorithm comparative evaluation, fairness audit across eight demographic subgroups, action-list stability analysis (Jaccard=0.670±0.065), and probability calibration assessment complete a methodological standard for workforce analytics models destined for APCS integration. Keywords: Employee attrition; data leakage; SMOTE; PR-AUC; cross-validation; imbalanced classification; APCS integration; fairness; Brier score; recall@k.
Сведения · Details
- DOI 10.5281/zenodo.22145250
- Журнал · Journal Central Asian Journal of STEM
- Язык · Language английский
- Цитирования (Crossref) · Citations 0