Explainable AI for employee turnover prediction: a SHAP-based intelligent analytics approach

Abstract

Employee turnover prediction is a critical challenge in human resource management. Existing studies emphasise predictive accuracy but generally treat interpretability as an afterthought, and applications of SHapley Additive exPlanations (SHAP) in this domain typically stop at single-level feature importance without reporting threshold calibration, subgroup fairness, or external validation. This study develops a four-layer SHAP-based explainable analytic protocol that systematically integrates global feature importance, feature effect, pairwise interaction, and local explanation with cohort analysis, with each layer attached to a specific managerial question, and that augments these four layers with threshold calibration, subgroup fairness audit, and cross-dataset external validation as standard reporting modules. Four machine learning models (XGBoost, Random Forest, LightGBM, Logistic Regression) are trained on the IBM HR Analytics benchmark under a leakage-corrected pipeline that places SMOTE inside cross-validation. XGBoost reaches a test-set ROC-AUC (Receiver Operating Characteristic Area Under the Curve) of 0.773 (95% confidence interval 0.716 to 0.828); it is not statistically superior to a Logistic Regression baseline and is selected as the SHAP base learner for its TreeSHAP compatibility and interaction-analysis capability rather than for predictive performance. The dummy-aggregated SHAP top 10 places EnvironmentSatisfaction, JobSatisfaction, StockOptionLevel, OverTime, and MonthlyIncome at the top of the ranking. Threshold sweep, fairness audit (which reveals a substantial age-band recall gap), and external validation on the Saudi Employee Attrition dataset show that the protocol transfers as a method while the specific feature rankings, the OverTime × MonthlyIncome interaction, and the dominant subgroup disparity are dataset-specific and require per-deployment validation. The protocol is positioned as a proof-of-concept analytic framework supported by an explicit responsible-AI discussion rather than as a substantive claim about which features drive turnover in general.