Employee Turnover Prediction: The impact of employee event features on interpretable machine learning methods

摘要

The purpose of this study is to study the impact of employee event features on classifier performance to detect employee likelihood to turnover meanwhile maintaining the interpretability of classifiers for benefits of retention interventions development. In order to maintain feature importance four machine learning methods has been selected on this study, consist of (1) Logistic Regressions Method; 2) Random Forest Method; (3) Gradient Boosting Tree Method and (4) Extreme gradient boosting method. The result show that employee event features significantly improve classifiers performance. The extreme gradient boosting method has the best performance with 98.03% Accuracy, 90.79% F1 Score, 97.18% Precision and 85.19% recall. Top three feature importance to predict employee turnover are Merit increase compare with market, Ratio of annual leave hours on average - last month and Ratio of annual leave hours on average - last 2 months respectively.