Federated Recommendation Systems (FRS) enable privacy-preserving model training by keeping user data on edge devices. However, the practical deployment of FRS in Edge-Cloud environments faces significant challenges due to system and statistical heterogeneity. Existing FRS participant selection strategies struggle to dynamically balance the trade-off between model convergence speed and recommendation quality in such volatile environments. To address this, we formulate the FRS participant selection problem as a normalized utility cost addressing the model quality and system efficiency. Next, we propose a dynamic participant selection framework incorporating a Multi-Armed Bandit (MAB)-based solver for multimodal FRS. We design a client-utility function that jointly evaluates historical Client Performance Reputation, data quality, and real-time system latency. By leveraging an Upper Confidence Bound strategy, our framework effectively balances the exploration of under-sampled clients with the exploitation of high-performing ones. We validate the proposed approach on a realistic edge-cloud testbed implementation using a multimodal movie-recommendation task. Experimental results demonstrate that our MAB-driven approach outperforms other baselines across eight different data-skew scenarios. Specifically, it improves training efficiency by 32-50% while improving model quality metrics such as Recall@50 by up to around 5%

翻译：联邦推荐系统通过将用户数据保留在边缘设备上，实现了保护隐私的模型训练。然而，在边缘-云环境中实际部署联邦推荐系统时，由于系统和统计异质性的存在，面临重大挑战。现有的联邦推荐系统参与者选择策略难以在此类动态环境中平衡模型收敛速度与推荐质量之间的权衡。为解决这一问题，我们将联邦推荐系统参与者选择问题形式化为一个归一化效用成本模型，用于衡量模型质量与系统效率。随后，我们提出了一种动态参与者选择框架，该框架集成了一种基于多臂赌博机的求解器，适用于多模态联邦推荐系统。我们设计了一个客户端效用函数，该函数联合评估历史客户端性能声誉、数据质量及实时系统延迟。通过采用上置信界策略，我们的框架有效平衡了对采样不足客户端的探索与高性能客户端的利用。我们在一个真实的边缘-云测试平台上，利用多模态电影推荐任务验证了所提方法。实验结果表明，在八种不同数据倾斜场景下，我们的多臂赌博机驱动方法优于其他基线方法。具体而言，该方法将训练效率提升了32-50%，同时将模型质量指标（如Recall@50）提升了约5%。