FedFast: Going Beyond Average for Faster Training of Federated Recommender Systems

Federated learning (FL) is quickly becoming the de facto standard for the distributed training of deep recommendation models, us-ing on-device user data and reducing server costs. In a typical FLprocess, a central server tasks end-users to train a shared recommen-dation model using their local data. The local models are trained over several rounds on the users’ devices and the server combinesthem into a global model, which is sent to the devices for the pur-pose of providing recommendations. Standard FL approaches userandomly selected users for training at each round, and simply average their local models to compute the global model. The resulting federated recommendation models require significant client effortto train and many communication rounds before they converge to asatisfactory accuracy. Users are left with poor quality recommendations until the late stages of training. We present a novel technique, FedFast, to accelerate distributed learning which achieves goodaccuracy for all users very early in the training process. We achievethis by sampling from a diverse set of participating clients in each training round and applying an active aggregation method that propagates the updated model to the other clients. Consequently, with FedFast the users benefit from far lower communication costsand more accurate models that can be consumed anytime during the training process even at the very early stages. We demonstrate the efficacy of our approach across a variety of benchmark datasetsand in comparison to state-of-the-art recommendation techniques