Comparative Analysis of Resampling and Feature Selection Methods for Employee Turnover Prediction

Abstract

Employees are the most valuable assets of each company. Unexpected employee turnover imposes something between %30 and %150 of the employee's annual salary to the company. In this study, different data balancing methods were applied to regulate the imbalances in the data set and to handle imbalanced data problem. In addition, to reduce the number of features in the data set, RFE and Boruta feature selection techniques were applied to compare their performance. We applied prediction algorithms from 3 different categories including classic machine learning, ensemble methods and deep learning. Overall, oversampling method has been shown to perform better than undersampling. Among the algorithms, XGBOOST achieved the highest performance with %90.90 F1 Score.