Time-Series Forecasting Based on Fuzzy Cognitive Visibility Graph and Weighted Multisubgraph Similarity
Yuntong Hu, Fuyuan Xiao
Abstract
This article aims to address the problem of time-series forecasting. Current state-of-the-art forecasting models lack the ability to mine the spatiotemporal dependence. How to mine more useable features of time series to make accuracy predictions is still an open issue. To address these challenges, from the perspective of fuzzy interaction between nodes, we propose a novel network constructing model called fuzzy cognitive visibility graph (FCVG) for time series to convert the time series into a pair of directed weighted graphs. To calculate the similarity between nodes in the FCVG, we develop the weighted multisubgraph similarity (WMSS). With these tools, we introduce the prediction based on fuzzy similarity distribution (PFSD), a novel forecasting method for time series, that can efficiently capture the spatiotemporal dependence. A time series is converted into a network by the FCVG, and the similarity scores between nodes are calculated through the WMSS. Based on the normalized similarity distribution, the predictions of time series are made. Extensive experiments on different datasets confirm the benefits of leveraging fuzzy interaction in time-series forecasting. Moreover, the construction cost index is predicted to show how to apply PFSD to forecast a specific time series.