A novel model based on graph kernel and S-R score in visibility graph for time series forecasting

Abstract

Time series contains rich historical information. Analyzing and utilizing this information to predict the future changes of the observed object has garnered widespread attention. The visibility graph method is an important branch in time series prediction. However, the approach of reducing interference and redundancy while leveraging the valid information of the visibility graph is void. Inspired by the idea of maximum relevance and minimum redundancy, we propose a Similarity-Redundancy (S-R) score to measure the contribution of different nodes after using graph kernel methods to calculate node similarities. Based on the proposed S-R score method, the selected high-quality nodes have both strong predictive ability (high correlation with the target node) and low information redundancy (low redundancy with other nodes). We conducted experiments on the proposed time series prediction model using the M-Competition datasets. The results show that the proposed S-R score can provide more accurate predictions.