Federated recommenders: methods, challenges and future.

Abstract Web users are flooded with information on the internet, and they feel overwhelmed by the different choices they have to make online daily. Recommender systems come to their rescue by suggesting products best aligned with their interests. To achieve this, traditional recommenders transfer users' personal data from the client to the server and dig for information about the user's interests and tastes. Moving data to the cloud violates the user confidentiality requirement and poses severe threats to user privacy and security. Moreover, with the tremendous increase in data size, it is no longer possible to collect and process massive data in the cloud. With the emergence of federated learning, numerous innovative recommender models are devised to solve these issues. In these models, the user data never leaves the client-side, and only the inferred results are sent back to the server for aggregating and updating the master model. Hence, the federated recommenders preserve user privacy and save the hassle of transferring enormous data to the cloud. This paper meticulously studies the recently proposed federated recommenders and classifies them based on the enhancements introduced in the prediction model, security scheme, or optimization technique. We identify the challenges faced by current federated recommenders and observe that most issues are inherently due to various aspects of federated learning, such as heterogeneous and non-IID data, malicious users, distributed framework, and non-reliable edge devices. While some emerge due to the coupling of the recommendation process in the federated paradigm. This research summarizes the current limitations, highlights the areas that need improvements, and presents future paths. In short, it paves the way for the development of robust federated recommenders that can handle the challenges of federated learning and, at the same time, generate high-quality recommendations. [ABSTRACT FROM AUTHOR]