Predicting Employee Attrition Using XGB Classifier

Abstract

Employee attrition is a critical problem in terms of cost and disruption of productivity. Therefore, it is important for organizations to predict which employees are likely to leave. The present paper will be premised on the XGBoost classifier that predicts attrition using the IBM HR Analytics data on Kaggle with 1,470 records of the employees and their demographic, job-related and performance data. The nominal variables were one-hot-coded and the label of the target transformed and stratified during the construction of a training set and a test set assisted in the preparation of the data. Hyper parameter optimization and over sampling techniques as well as feature engineering were chosen to ensure the optimization of the model to deal with the imbalance of the classes. The overall model of the XGBoost was 87.76 and this was reasonable in classifying the employees who remained and those who lapsed. The ‘Over Time’, ‘Monthly Income’ and the ‘Job Satisfaction’ are some of the factors that resulted into high level of impact on attrition. This paper has identified the merits and demerits of machine learning in HR analytics and has uncovered ethical concerns of fairness, transparency and privacy security of employees as applied to the use of predictive models to control the human resource.