MHVG2MTS: Multilayer Horizontal Visibility Graphs for Multivariate Time Series Analysis

Abstract: Understanding the properties of time-indexed multivariate data has been a predominant topic mainly to address open issues in multivariate time series analysis. Usually, the methodologies used to analyze multivariate time series are based on adapting approaches for univariate settings or on assumptions and parameters for specific problems. A different strategy uses complex network to obtain an additional and reduced representation of temporal and causal properties of the time series data. Recent strategies involve mapping multivariate time series into high-level network structures, specifically into multiplex networks representing interconnections between contemporary timestamps of different time series components. In this work, we propose a new mapping method that takes advantage of the entire structure of multilayer networks. We introduce the multilayer horizontal visibility graph that is based on the new concept of cross-horizontal visibility between lagged timestamps of different components, which allows describing the cross-dimension dependencies via inter-layer edges. We use a set of existing topological measures of multilayer networks as well as a novel measure to evaluate and validate our approach, which is parameter-free, does not require data pre-processing and is applicable to any kind of multivariate time series data. We provide an extensive experimental evaluation, where we explore the proposed topological measures, showing that the inter-layer edges based on cross-horizontal visibility preserve more information about the time series data after the mappings, information that would inevitably be lost using mapping methods that result in single-layer and multiplex structures. We also verify that the information mapped by the inter-layer edges is not enough on its own, but that it complements the data information captured by the commonly used intra-layer edges.