GWO-RF: A Grey Wolf Optimized Random Forest Model for Predicting Employee Turnover

Abstract

This study proposes an employee turnover prediction model (GWO-RF) that combines Grey Wolf Optimization (GWO) algorithm with Improved Random Forest (LPRF). The model optimizes node splitting strategy by combining C4.5 information gain rate and CART Gini coefficient (constraint condition α+β=1) through linear programming. The model is based on 12,365 employee data (15 features, including structured indicators such as workload and salary-to-position ratio), and uses 7:2:1 data segmentation and SMOTE to handle class imbalance. Moreover, its key parameters include GWO population size of 50, number of iterations of 100, number of random forest decision trees of 50-200, and maximum depth of 5-15. The test set results show that the model has an AUC of 0.923±0.008 and an F1-score of 0.871. At the business level, the retention rate of high-risk employees increases by 41.9% (p<0.01), and the cost of single intervention decreases by 54.3%. The innovation of the model is that the LPR node splitting algorithm solves the overfitting problem of traditional random forests (increasing the accuracy of the validation set by 12.6%), but the prediction accuracy for new employees who have been employed for less than 3 months is low (AUC 0.782). Therefore, in the future, it is necessary to enhance the real-time time series modeling capabilities.