Network self attention for forecasting time series.

Recently, attention mechanism has become a hot research topic. Its ability to assign global dependencies from input to output makes it widely used. Meanwhile, although there are some forecasting methods for time series, how to mine more information of time series and make more accurate predictions of it is still an open question. To address this issue, we propose network self attention to mine more information of time series. And we propose a novel forecasting model for time series, in which the similarity scores are learned by a recurrent neural network (RNN) based on network self attention. The similarity scores of nodes in network that is converted from time series by visibility algorithm are learned by RNN at the first step. Afterwards, the final network attention value is calculated. Finally, the prediction of time series is made with the final network attention value. To test the ability of our method to forecast time series, we make predictions of construction cost index (CCI), M1 and M3 datasets. Experiment results indicate that our method can make better predictions for some time series compared to other methods. Meanwhile, robustness test indicates that our method is of good robustness. (c) 2022 Elsevier B.V. All rights reserved.

Automated summary

This study proposes a novel forecasting model for time series based on network self attention, which aims to mine more information from time series and improve prediction accuracy. Experimental results demonstrate that the proposed method outperforms other methods in predicting construction cost index, M1, and M3 datasets, and exhibits good robustness.