Communication-efficient distributed online prediction by dynamic model synchronization

We present the first protocol for distributed online prediction that aims to minimize online prediction loss and network communication at the same time. This protocol can be applied wherever a prediction-based service must be provided timely for each data point of a multitude of high frequency data streams, each of which is observed at a local node of some distributed system. Exemplary applications include social content recommendation and algorithmic trading. The challenge is to balance the joint predictive performance of the nodes by exchanging information between them, while not letting communication overhead deteriorate the responsiveness of the service. Technically, the proposed protocol is based on controlling the variance of the local models in a decentralized way. This approach retains the asymptotic optimal regret of previous algorithms. At the same time, it allows to substantially reduce network communication, and, in contrast to previous approaches, it remains applicable when the data is non-stationary and shows rapid concept drift. We demonstrate empirically that the protocol is able to hold up a high predictive performance using only a fraction of the communication required by benchmark methods.