<p>Matrix factorization (MF) is a prevalent technique in recommender systems, its reliance on centralized storage raises significant privacy concerns. Federated learning (FL), as a distributed learning framework, has emerged as a promising solution. However, even in FL, the transmission of models or gradients can still pose risks to user privacy. To address this, we propose a privacy-preserving semi-decentralized matrix factorization (PSDMF) framework for personalized rating prediction, incorporating privacy-preserving grouping and random transmission to enhance privacy. PSDMF combines the advantages of decentralized and client-server architectures from FL, allowing users to directly exchange model parameters within groups for training while preserving the server’s aggregation role. Our framework comprises three key phases. Firstly, we introduce a privacy-preserving grouping approach to identify communication neighbors with similar preferences for users. Secondly, we employ a random transmission strategy to enable each group to build a group item latent matrix tailored to specific user group preferences, without compromising privacy. Finally, the server collects and aggregates all the group models to generate a global latent matrix that captures broader preferences. Experimental validation conducted on five real-world movie rating datasets demonstrates that PSDMF ensures better recommendation performance while preserving user privacy.</p>