Federated Learning Based Framework for Privacy-Preserving Employee Attrition Prediction in Human Resource Analytics

Abstract

Employee attrition represents a critical challenge for contemporary organizations, impacting productivity, elevating recruitment costs, and eroding institutional knowledge. While machine learning offers robust tools for predicting turnover, its application is frequently hindered by “small data” constraints within individual firms and the sensitive nature of human resource (HR) records, which complicates collaborative data-sharing. This paper introduces FedHR, a federated learning framework designed to enable multi-institutional collaboration without the exchange of raw data. In FedHR, predictive models are trained locally within participating organizations, with only model parameters aggregated centrally. This architecture ensures compliance with rigorous privacy standards like GDPR and CCPA, further reinforced by differential privacy mechanisms. Using a simulated consortium of four diverse enterprises, we demonstrate that FedHR achieves an F1-score of 0.88, significantly outperforming isolated local models (0.74) and approaching the efficiency of centralized baselines (0.90). These results establish federated learning as an ethically responsible and technically viable infrastructure for strategic talent analytics.