Survey of recommender systems based on federated learning
锋 梁, 恩跃 羊, 微科 潘, 强 杨, 仲 明 · Scientia Sinica Informationis · 2021
With the development of the Internet and mobile computing, people's online behaviors have generated increasing amounts of data.In order to select items that users may like from massive data, recommender systems are indispensable.However, traditional recommendation algorithms need to collect user data to the server to build the model, which will leak user privacy.Recently, Google has proposed a new learning paradigm called federated learning for machine learning problems that require user data to be collected for modeling.The combination of federated learning and recommender systems enables federated recommendation algorithms to always keep user data in clients during the modeling process, so as to protect user privacy.In this study, the research works on the combination of federated learning with recommendation algorithms are surveyed. Then, the research development on federated recommendation algorithms is analyzed from three perspectives, namely, design of architectures, federalization of models, and application of privacy-preserving technology.Finally, some research directions and prospects for recommender systems based on federated learning are discussed.