Subwindow-based Topological Features to Deep Learning Models for Time Series Forecasting

Abstract

In time series forecasting, learning the pattern that underlies on the previous time steps is important to predict the upcoming time steps accurately. Depending on the embedding window dimension of the previous observations, the pattern of the time series can be captured either in a local scope or global scope. However, in many cases, capturing the pattern in both scopes are necessary in order to make the prediction more accurate, so it remains as a challenging task. In this paper, Subwindow Topology is proposed to effectively capture the pattern of the time series in both local and global scope. The proposed method extracts the topological features from each subwindow of the time series in order to capture the local pattern and integrates the topological features with the input features of the time series. The integration of the input features and topological features is utilized as the input data by window embedding for prediction so that the global pattern can also be captured while preserving the information about the local pattern. The proposed method is evaluated by comparing the performance with other forecasting models using public time series datasets of various domains and measuring the boosting effect to examine whether it can improve the forecasting performance when applied to various forecasting models. Also, ablation study is conducted to investigate the necessity of the components of the proposed method for effectively capturing the local and global pattern of the time series.