Federated Learning with Variational Autoencoder for Popularity Profile Prediction

Motivated by increasingly exploding data traffic of online video services, the prediction of the popularity profile of video contents becomes very important for network traffic prediction, recommendation systems, and wireless caching. This paper proposes a federated learning-based popularity prediction scheme using a variational autoencoder (VAE), which copes with the situation where users are moving and/or their data privacy should be protected. Users are participants of federated learning, and the VAE model is trained by user's own request history; afterwards, randomly generated samples from the pretrained decoder of VAE can mimic the original popularity profile. We adopt the MovieLens dataset to validate the proposed model, and experimental results show that our scheme predicts the popularity profile almost perfectly.