A scalable framework for large time series prediction

Knowledge discovery systems are nowadays supposed to store and process very large data. When working with big time series, multivariate prediction becomes more and more complicated because the use of all the variables does not allow to have the most accurate predictions and poses certain problems for classical prediction models. In this article, we present a scalable prediction process for large time series prediction, including a new algorithm for identifying time series predictors, which analyses the dependencies between time series using the mutual reinforcement principle between Hubs and Authorities of the Hits (Hyperlink-Induced Topic Search) algorithm. The proposed framework is evaluated on 3 real datasets. The results show that the best predictions are obtained using a very small number of predictors compared to the initial number of variables. The proposed feature selection algorithm shows promising results compared to widely known algorithms, such as the classic and the kernel principle component analysis, factor analysis, and the fast correlation-based filter method, and improves the prediction accuracy of many time series of the used datasets.