Privacy-preserving cross-network service recommendation via federated learning of unified user representations

Abstract

With the emergence of cloud computing, the Internet of Things, and other large-scale environments, recommender systems have been faced with several issues, mainly (i) the distribution of user–item data across multiple information networks, (ii) privacy restrictions and the partial profiling of users and items caused by this distribution, (iii) the heterogeneity of user–item knowledge in different information networks. Furthermore, most approaches perform recommendations based on a single source of information, and do not handle the partial representation of users’ and items’ information in a federated way. Such isolated and non-collaborative behavior, in multi-source and cross-network information settings, often results in inaccurate and low-quality recommendations. To address these issues, we exploit the strengths of network representation learning and federated learning to propose a service recommendation approach in smart service networks. While NRL is employed to learn rich representations of entities (e.g., users, services, IoT objects), federated learning helps collaboratively infer a unified profile of users and items, based on the concept of anchor user, which are bridge entities connecting multiple information networks. These unified profiles are, finally, fed into a federated recommendation algorithm to select the top-rated services. Using a scenario from the smart healthcare context, the proposed approach was developed and validated on a multiplex information network built from real-world electronic medical records (157 diseases, 491 symptoms, 273 174 patients, treatments and anchors data). Experimental results under varied federated settings demonstrated the utility of cross-client knowledge (i.e. anchor links) and the collaborative reconstruction of composite embeddings (i.e. user representations) for improving recommendation accuracy. In terms of RMSE@K and MAE@K, our approach achieved an improvement of 54.41% compared to traditional single-network recommendation, as long as the federation and communication scale increased. Moreover, the gap with four federated approaches has reached 19.83 %, highlighting our approach’s ability to map local embeddings (i.e. user’s partial representations) into a complete view.