Predicting Employee Attrition Using Ensemble Machine Learning Techniques: A Comparative Study

Abstract

One of the most important problems to solve for any organization is employee attrition because of the effect on workforce stability and loss of knowledge and experience. Historical HR analytics solutions are likely to be reactionary and might not be able to identify complex relationships in employee data. Based on this work, a soft voting scheme is proposed for the ensemble prediction of employee attrition with high accuracy. Thee research consists of the IBM HR Analytics dataset, containing 1,470 employee records and 35 features, from the Kaggle database. Synthetic Minority Over-Sampling Technique (SMOTE) and Random Over-Sampling (ROS) methods are used to tackle the class imbalance problem. The proposed ensemble is a combination of Extreme Gradient Boosting (XGBoost), Random Forest and calibrated Logistic Regression classifiers. The experimental results indicate that the accuracy of the model is 97.72% and the F1 score is 97.74%. The results prove that data-driven predictive systems can play a key role in implementing proactive employee retention strategies.