CoxRF: Employee Turnover Prediction Based on Survival Analysis

In human resource management, employee turnover prediction is very important for company operation since the leave of key employees can bring great loss to companies. However, most existing researches focused on employee-centered turnover prediction, while ignored the historical events of turnover behaviors as well as the longitudinal data of each work. Therefore, in this paper we propose a turnover prediction algorithm named CoxRF from an event-centered perspective, which combines the statistical results of survival analysis with ensemble learning and therefore simplifies the problem to a traditional supervised binary classification problem. In addition, we come up with the concepts of "event-person" and "time-event" to help construct survival data with censored data. We compare CoxRF with several baseline methods on a real dataset crawled from the biggest professional social platform of China. The results demonstrate its effectiveness for predicting turnovers. Besides, we have the following findings: i) gender plays a key role in employee turnover behavior and female employees have a higher turnover rate than male employees; ii) the external environmental factor such as GDP growth also contributes greatly to employee turnovers, which is hardly considered in previous data-driven studies; iii) The employee turnover behavior varies across different industries, where the IT group has significantly higher turnover rate than the government group; iv) Employees with good education backgrounds have a higher turnover rate than those with ordinary backgrounds after working for 3 to 5 years.