An ML-Driven Participant Selection Technique for Federated Recommendation System in Edge-Cloud Computing

Abstract: Recommendation systems (RS) personalize content by analyzing user preferences, but typically require centralized collection of user data, raising privacy and scalability concerns. Federated Recommendation Systems (FRS) address these issues by enabling distributed, privacy-preserving model training across edge devices, keeping raw data on-device. Although existing FRS frameworks benefit from on-device feature extraction and privacy preservation, they suffer from heterogeneous device capabilities, non-independent and identically distributed (non-IID) data, and communication bottlenecks. To overcome these limitations, we propose a multi-objective reinforcement learning (RL) participant selection that jointly optimizes historical client performance reputation (CPR), data utility, and system efficiency. First, we define a composite client-utility function combining CPR, system capability, and data quality. Next, we embed this utility into a multi-armed bandit (MAB) framework and dynamically balance exploration-exploitation to select participants. Finally, we practically implement our approach using the PySyft framework on an edge-cloud testbed, and evaluate it on a multimodal movie-recommendation task built from the MovieLens-100K dataset. Across four different skewed data-partition scenarios, our MAB-based selection accelerates convergence by 32-50% in time-to-target AUC and reduces total wall-clock training time by up to 46%, while matching or slightly improving final AUC, NDCG@50, and Recall@50 compared to existing FRS baselines. Our results demonstrate that adaptive, reward-driven client sampling can substantially enhance both efficiency and fairness in real-world federated deployments.