Abstract
Reinforcement learning methods have been used to optimize long-term userengagement in recommendation systems. However, existing reinforcementlearning-based recommendation systems do not fully exploit the relevance ofindividual user behavior across different platforms. One potential solution isto aggregate data from various platforms in a centralized location and use theaggregated data for training. However, this approach raises economic and legalconcerns, including increased communication costs and potential threats to userprivacy. To address these challenges, we propose \textbf{FedSlate}, a federatedreinforcement learning recommendation algorithm that effectively utilizesinformation that is prohibited from being shared at a legal level. We employthe SlateQ algorithm to assist FedSlate in learning users' long-term behaviorand evaluating the value of recommended content. We extend the existingapplication scope of recommendation systems from single-user single-platform tosingle-user multi-platform and address cross-platform learning challenges byintroducing federated learning. We use RecSim to construct a simulationenvironment for evaluating FedSlate and compare its performance withstate-of-the-art benchmark recommendation models. Experimental resultsdemonstrate the superior effects of FedSlate over baseline methods in variousenvironmental settings, and FedSlate facilitates the learning of recommendationstrategies in scenarios where baseline methods are completely inapplicable.Code is available at \textit{https://github.com/TianYaDY/FedSlate}.