Federated Learning for Cold Start Recommendations

With the increasing desire to protect the users’ data and provide data privacy, few recent research works have been focusing on federated learning to avoid centralized training for various machine learning problems. Federated learning involving clients such as mobile devices has been widely explored and there is little exploration done toward training a federated model involving corporate companies or organizations. In this work, we want to understand the capability of a federated recommendation model using three news publishers’ users’ data to generate cold-start recommendations. We train a recommendation model using federated learning from three news publishers and evaluate the cold-start recommendation performance of our federated model against a model trained with each news publisher’s data alone. Our results demonstrate that federated learning boosts the cold start recommendation for all news publishers and there is a higher ranking performance when federated with news publishers that have similar user behavior.