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

Area of Science:

• Human Resource Management
• Explainable Artificial Intelligence (XAI)
• Machine Learning

Background:

• Employee turnover prediction is crucial for HR, but existing methods often lack interpretability.
• SHapley Additive exPlanations (SHAP) applications typically focus on basic feature importance, omitting critical validation steps.
• There's a need for a comprehensive, interpretable framework for employee turnover analysis.

Purpose of the Study:

• To develop and validate a four-layer SHAP-based explainable analytic protocol for employee turnover prediction.
• To integrate global feature importance, feature effect, interaction analysis, and local explanations with cohort analysis.
• To augment the protocol with threshold calibration, subgroup fairness audits, and external validation.

Main Methods:

• Developed a four-layer SHAP protocol integrating various explanation techniques and responsible AI modules.
• Trained four machine learning models (XGBoost, Random Forest, LightGBM, Logistic Regression) on the IBM HR Analytics benchmark using a leakage-corrected pipeline.
• Performed threshold calibration, subgroup fairness audit, and external validation on the Saudi Employee Attrition dataset.

Main Results:

• The SHAP protocol demonstrated transferability as a method across datasets.
• XGBoost achieved an ROC-AUC of 0.773, but was not statistically superior to a Logistic Regression baseline.
• Fairness audit revealed a substantial age-band recall gap; specific feature rankings and interactions were dataset-dependent.

Conclusions:

• The developed SHAP protocol offers a robust framework for interpretable employee turnover prediction.
• While the protocol methodology is transferable, specific findings require per-deployment validation due to dataset specificity.
• Emphasizes the importance of responsible AI considerations, including fairness and interpretability, in HR analytics.