Implementing Feature-ranking Techniques for Predicting Employee Attrition Through an Optimised XGBoost Classifier

Abstract

Employee attrition has become a critical concern for all organisations due to its detrimental impact on workplace productivity and the timely achievement of organisational objectives. The consequences of employee attrition include disruption in workflow and team dynamics. Moreover, it also consists of the expenses associated with re-employment and retraining. Thus, proactively identifying and addressing factors contributing to attrition becomes essential. Organisations are progressively moving towards machine learning (ML) techniques for predicting employee attrition to accomplish this purpose. However, predicting and understanding the reasons behind employee attrition poses a challenge for HR managers. Therefore, this study delves into the realm of employee attrition by proposing an optimised ensemble classifier XGBoost to predict the possibility of employee attrition, aiming to enable proactive talent management, a departure from the traditionally reactive approach. This study presents a feature-based ranking methodology to identify the features behind employee attrition. At first, three distinct feature-ranking methods are employed to rank the features, notably Chi-square, ReliefF and mRMR. The results show that ReliefF identified the features more effectively than others. To this end, the top 25 features ranked through ReliefF predict employee attrition at an accuracy of 93.10%, eventually reducing certain ineffectual features.