Employee Relationship Network Modeling and Turnover Prediction Based on Graph Neural Networks

Employee turnover poses persistent challenges for organizations, disturbing operational continuity and strategic planning. Traditional machine learning models, while effective to a certain degree, often fail to capture relational and temporal patterns inherent in workforce dynamics. This study presents a graph-based simulation framework that models employee relationships as a network and applies Graph Neural Network (GNN) architectures, including Graph Attention Networks (GANs), graph sample and aggregate (namely GraphSAGE), and Graph Convolutional Networks (GCNs), to learn interdependencies influencing attrition. A stacked ensemble approach integrates these models using logistic regression to enhance predictive robustness. Evaluated on the IBM HR Analytics dataset, the proposed framework achieves 93.2% accuracy. To support managerial decision-making, explainable methods are used to identify critical risk factors such as job satisfaction, workload, and environmental conditions. This process-oriented modeling approach simulates employee attrition dynamics and generates interpretable outputs that guide proactive retention strategies.