Explainable and Fair Employee Attrition Prediction: A Cost-Sensitive Hr Analytics Approach

Abstract

Organisations increasingly use predictive HR analytics to identify employees at risk of leaving. Much of this work, however, reports only classification accuracy, neglecting whether complex models outperform simple ones, what drives their predictions, how predictions should be converted into decisions, and whether they treat employee groups fairly. This study addresses these four questions within a single, transparent analytical framework. Using the public IBM HR Analytics Employee Attrition dataset (N = 1,470; 30 predictors; 16.1% attrition), eight classifiers were compared through 5-fold cross-validation repeated three times, with differences tested by the corrected resampled t-test and the Friedman test. Logistic regression, random forest and gradient boosting were then tuned and evaluated on a stratified hold-out set of 441 employees. Regularised logistic regression achieved a cross-validated AUC of 0.835, statistically outperforming random forest (0.799), support vector machine (0.806), decision tree, k-nearest neighbours and naive Bayes, and matching gradient boosting (0.806 untuned; 0.836 tuned). On the hold-out set, the selected class-weighted logistic model achieved an AUC of 0.817 (95% CI 0.758–0.874) and captured 43.7% of leavers within the top 10% of risk scores. Overtime was the dominant predictor across all models, followed by years since promotion, job-hopping history, job satisfaction and business travel. A cost-optimal threshold of 0.62 reduced expected attrition cost by 12.8% relative to no intervention under base assumptions. The fairness audit revealed that employees under 30 and single employees were flagged at substantially higher rates and that true-positive rates differed by up to 22 percentage points across age groups. The study shows that interpretable models can match complex machine learning in HR prediction, and it offers an end-to-end template linking prediction, explanation, economics and fairness for responsible people analytics.