The Taub Faculty of Computer Science Events and Talks

Multivariate time series forecasting differs from univariate time series forecasting by trying to model the dependencies between the different time series in order to make a more precise forecasting. Despite reaching better results in a multivariate setting, classical and deep learning multivariate models are not scalable, having their total number of parameters growing square of the number of time-series. We present in this paper a novel paradigm in how deep learning models shall be adapted to high dimension multivariate time series forecasting. Our paradigm adapts any deep learning model into a scalable reduced and distributed model, focusing on the parameters necessary to capture the essential dependencies of the data. We leverage knowledge from the most meaningful dependencies with a cluster based decomposition of the data while still learning from global dependencies between all the time series via a small global component. A third component efficiently model the common trends of the different timeseries via a univariate model. Our model brings state of the art results on smaller datasets and mainly on bigger ones while being more efficient memory wise and time wise.