Research On Employee Turnover Prediction Before And After Pandemic Based On Model Fusion And Oversampling

With the rapid development of market economy and the diversification of social concepts,the turnover intention of enterprise employees has increased compared with the past,and the human resource management problems brought by employee turnover are increasingly prominent.The Human Resources Department needs to evaluate the probability of employees turnover,understand the reasons for employees turnover,and give a plan to retain employees accordingly.In previous studies,most of the employee turnover problems lacked the processing of unbalanced data,and relied too much on the classification of a single model,which led to poor prediction results for a small number of employees who left.Therefore,this paper will use the oversampling method to balance the data set,and use the linear weighted voting method to fuse the model to improve the classification efficiency.At the same time,this paper will also try to adjust the critical value of prediction probability and the fusion model for some samples to enhance the prediction ability of the model for minority category.This paper mainly uses RF and XGBoost as the base learners,selects two sets of employee turnover data before and after the pandemic,and focuses on the enhancement effect of oversampling method,adjusting the critical value of prediction probability and model fusion method on the comprehensive classification performance of the model.After data preprocessing,this paper successively uses SMTOE method,adjusting critical value,model fusion method,and verifies the effectiveness of these methods.By comparing the importance of variables output from the statistical model,the changes of the factors affecting employee turnover before and after the pandemic were analyzed and compared.The empirical research shows that the oversampling method and adjusting the critical value of prediction probability are helpful to improve the comprehensive classification efficiency of the model,especially to improve the prediction ability of minority category.The effect of the two model fusion and multi model fusion methods is also better than that of the single model,which also shows that the linear weighted voting fusion model is a simple and feasible optimization method.