ComESN: Compressive Echo State Network for Multi-task Time-series Prediction

Time-series prediction has been widely studied for single-variable applications, but shows poor performance for multi-variable applications due to the complex correlation. Multivariable prediction is often transferred into many single-variable tasks, and results into high computational cost. The prediction performance decreases exponentially as the dimensions increase. The compression of multiple prediction tasks is necessary to provide a trade-off between computational cost and prediction accuracy. For this purpose, compressive echo state neural network (ComESN) is proposed, which consists of a compressive module and a multi-task echo state network (MT-ESN). In this study, the locally linear embedding algorithm is adopted as compression method to reduce the number of prediction tasks for improving the computational efficiency. A novel MT-ESN is developed to further improve the performance of ESN for multi-task prediction. Due to highly efficient compression and unveiling correlation of multiple tasks, ComESN could cut down the computational cost while providing a promising prediction accuracy. Experiments on three real-world datasets illustrate that the proposed ComESN could compete with state-of-the-art methods with regard to computational cost and accuracy.