Explainable attrition risk scoring for managerial retention decisions in human resource analytics

Area of Science:

• Data Science
• Human Resources Management
• Organizational Behavior

Background:

• Employee turnover presents a significant operational challenge for organizations.
• Predictive analytics offers a potential solution for HR managers to forecast and mitigate turnover.
• Limited research exists on the interpretability and reliability of managerial forecasts in this domain.

Purpose of the Study:

• To develop and evaluate an interpretable predictive model for employee attrition.
• To provide HR managers with data-driven insights for proactive workforce management.
• To investigate the impact of model calibration on prediction accuracy.

Main Methods:

• Utilized the Employee Attrition dataset with pre-processing techniques (label encoding, feature scaling, SMOTE).
• Trained and optimized machine learning models (Random Forest) using grid search and cross-validation.
• Applied sigmoid calibration, LIME, permutation feature importance, and SHAP for interpretability and risk assessment.

Main Results:

• The Random Forest classifier achieved a high AUC-ROC score of 97.37%.
• Model calibration significantly reduced the Brier Score, indicating improved probability accuracy.
• Visualizations identified employees with the highest attrition risk.

Conclusions:

• A calibrated, interpretable, and risk-stratified model enhances HR decision-making for retention strategies.
• This data-driven framework facilitates a shift from reactive to proactive workforce management.
• Leveraging predictive analytics empowers HR leaders to improve employee retention and operational efficiency.