SAJSET-01-2023-0014 Investigating the role of negative sampling algorithm for employee turnover prediction using improved dynamic bipartite graph embedding

Abstract

Employee turnover has become a big challenge for organizations, specifically in the present competitive environment where experienced employees are the biggest asset of an organizations. Depending on how difficult it is to get the equivalent replacement, the cost of an experienced employee voluntary turnover is ranging from 1.5 to 5 times the employee’s annual salary. To some extent such losses affect organizational processes and alter the efficiency of their targeted plan. In an attempt to address the problem, previous researches centered on examining the impact factors such as employees and organization’s attributes. This work emphasizes on modelling employee’s time- related historical job experience as a dynamic bipartite graph to study a vector representation of employees and organizations. This is achieved by developing a model that generates a sequence for each vertex in the bipartite graph using a horary random walk (HRW) method and input the sequence to a skip-gram with negative sampling (SGNS) to get the vector representation for each vertex. We then combine the vectors with the employee’s basic information as input to machine learning classifiers to predict the employee’s turnover. The data was collected from one of China’s recruitment platform. Precisely, this study proposed a low complexity model called Improved Dynamic Bipartite Graph Embedding (IDBGE). Furthermore, the experimental results shows that our technique had significantly improved the performance of the employee turnover prediction on recall, F1, AUC_ROC with 4.24%, 4.43%, and 2.82% respectively.