Dynamic User Scheduling in Wireless Multimodal Federated Learning via Comprehensive Semantic Evaluation

With the rapid development of the Internet of Things (IoT), multimodal data is becoming increasingly abundant, bringing new opportunities for federated learning (FL). However, research on user scheduling for multimodal FL over wireless networks is still lacking and faces challenges of user heterogeneity and high communication latency. Existing scheduling methods often neglect the intrinsic semantic value of usergenerated data. In response to this, we propose a dynamic user selection strategy driven by comprehensive semantic-value evaluation. This strategy designs a comprehensive scoring mechanism to evaluate user contributions, which integrates semantic coverage, semantic gain, and gradient norm. We embed this mechanism within an online learning optimization framework to guide user selection-striking a balance among semantic value, communication efficiency, and long-term fairness. The goal of this method is to optimize the total training delay and final model performance. Simulation experiments show that compared with random selection and polling scheduling baselines, the proposed strategy can significantly accelerate the convergence speed of the model, improve the final model accuracy, and effectively reduce the overall training time.