A fast feature selection method in polynomial echo state networks

Recently, a polynomial echo state network (PESN), an extension of the original ESN, was proposed. Its polynomial output weights were augmented by the high order information of full input features. However, there may be noisy and redundant features to construct the polynomial function as the output weights, which results in the high computational complexity and degrade the testing accuracy. To eliminate the insignificant features and reduce computational burden, a fast feature selection method in polynomial ESN (FS-PESN) is presented. Firstly, a feature selection criterion is utilized to choose the appropriate features to construct the p-order reduced polynomial function as the output weights in PESN. Then, an iterative strategy is given to reduce the training computational burden in FS-PESN. Finally, ten benchmark regression data sets experiments are done which show that the proposed approach can obtain better prediction accuracy and less testing time than the PESN.