Explainable Artificial Intelligence for Employee Attrition Prediction: Integrating Predictive Performance with Actionable HR Retention Strategies

Abstract

Employee attrition is one of the most critical phenomena for contemporary companies as it directly impacts the bottom line. Although the application of machine learning (ML) methods for attrition prediction is a widespread practice, the task of providing these black-box models with any practical interpretability for end-users (e.g., human resources (HR) managers) remains underexplored. This paper aims to design an integrated framework that (i) identifies the best-performing ML classifiers for the given task and (ii) employs model-agnostic techniques for feature importance explanation. An extensive comparative performance analysis was carried out on the IBM HR Analytics Employee Attrition Dataset (N = 1470, 31 retained features). Six supervised classification tasks were implemented and evaluated using an 80/20 stratified train/test split and 5-fold stratified cross-validation. The best overall performance was reported by the gradient-boosted tree ensemble (held-out ROC-AUC = 0.812, accuracy = 86.1%), but class-weighted logistic regression demonstrated the best stability across cross-validation folds (mean ROC-AUC = 0.827) and significantly better minority-class recall (61.7%). In line with permutation importance analysis, partial dependence plots, and tree-specific interpretation, logistic regression coefficients identified the vital predictors of employee attrition, namely, work-life balance, stock options, satisfaction, time since the last promotion, business travel, and job role. The patterns discovered were incorporated into a conceptual framework, which included five specific actions for human resource managers to consider in order to improve employee retention. The findings were evaluated through the lens of evidence-based business practices, thus, demonstrating how the combined use of ML and XAI approaches could actualize a predictive model in practice.