Explainable Artificial Intelligence for Intelligent Human Resource Analytics: Predicting Employee Retention and Workforce Performance

Abstract

While employee retention and workforce performance are pivotal to the continuity of organizations, there may be less clarity around the factors influencing the output of employees in predictive HR models. This study developed and explained machine-learning models for employee retention and workforce performance and assessed gender-based differences in retention-model performance. Two independent cross-sectional datasets comprising 1,191 employees for retention classification and 368 employees for performance prediction were analyzed. Logistic regression and random forest classification were evaluated for retention, while multiple linear regression, random forest regression, and a mean baseline were compared for performance. Models were validated using repeated five-fold cross-validation and out-of-fold predictions. Permutation importance and SHAP were applied for model explanation, followed by sensitivity and gender-subgroup assessments. The retention random forest achieved an ROC-AUC of 0.824 and balanced accuracy of 0.735. Years of experience was the strongest retention predictor, followed by job stability, deserved promotion, emotional commitment, and training participation. The performance of the random forest achieved an R-squared of 0.686 and a root mean squared error of 0.393. Employee readiness was the dominant performance predictor, followed by organizational support culture, innovative HR strategies, and digital HR technologies. Gender-specific ROC-AUC and balanced accuracy were similar, although predicted retention rates and class-specific errors differed. Explainable machine learning can complement workforce analytics with beneficial predictability and interpretable individual employee factors to enable transparent workforce analytics. There's a need for human involvement, external validation, and ongoing fairness monitoring in the responsible use of HR.