Popularity Prediction with Federated Learning for Proactive Caching at Wireless Edge

File popularity prediction plays an important role in proactive edge caching. The widely-used methods for popularity prediction are based on centralized learning, which needs to collect the request information and even more personal information from users, incurring the privacy-disclosure risk. In this paper, we propose a method of predicting file popularity with federated learning to address the privacy issue, where the request data of each user for each file is only employed for local training at each user. To facilitate the popularity prediction at the base station (BS) without information disclosure as well as the supervised training of a neural network at each user, we let each user upload a weighted sum of its own preference and file popularity to the BS. The neural network is employed to predict the weighted sum at each user by training with the user’s historical request records for each file. The local models and uploaded results of the users are aggregated at the BS. We show the convergence of the proposed popularity prediction method with a synthetic dataset. We use simulation results with a real dataset to show that the proposed method performs closely to the centralized learning based method in terms of caching performance.