Robust Causal Inference for Irregularly Sampled Time Series from Dynamical Systems

0570099 - ÚI 2023 GB eng A - Abstrakt
Kathpalia, Aditi - Manshour, Pouya - Paluš, Milan
Robust Causal Inference for Irregularly Sampled Time Series from Dynamical Systems.
Dynamics Days Europe 2022. Abstract Book. Aberdeen: University of Aberdeen, 2022. s. 125-125.
[Dynamics Days Europe 2022. 22.08.2022-26.08.2022, Aberdeen]
Grant CEP: GA ČR(CZ) GA19-16066S
Grant ostatní: AV ČR(CZ) AP1901
Program: Akademická prémie - Praemium Academiae
Institucionální podpora: RVO:67985807
https://www.abdn.ac.uk/events/documents/Dynamics%20Days%20Abstract%20Book%202022%20online%20version%20(002).pdf

While equation-based approaches allow us to describe and model different dynamical systems, their limited domain of applicability and validity, especially in nonlinear and non-equilibrium circumstances, has led numerous data-driven approaches to come into play. Among these, causal inference from time series has proved itself as a useful tool for studying the interactions between coupled systems and giving insights into their underlying mechanisms. Although a lot of causality estimation techniques have been proposed for non-linear systems, they usually give spurious results when applied to time series with short length, missing samples and unevenly sampled data. CompressionComplexity Causality (CCC) [1] is a recently proposed causality measure inspired from complexity estimators based on lossless ‘data-compression’ algorithms and has been found to be robust to the above-mentioned limitations. However, this measure is still limited to scalar time series. To extend CCC applicability to complex real-world systems, which are most-often multi-dimensional, we propose a method that first symbolizes the time-series from an observable of a multidimensional dynamical system based on Taken’s method of time-delayed embedding. Time-delayed vectors are transformed into a one-dimensional sequence using permutation or ordinal patterns coding [2]. The combination of permutation coding and CCC enables us to propose and apply the novel ‘Permutation CCC (PCCC)’ on simulated data. Simulation analyses reveal that PCCC retains the original strengths of CCC and performs much better than some existing state-of-the-art approaches. We apply PCCC to some real-world data from climatology and paleoclimatology with missing samples, irregular sampling and/or short length to make useful inferences about the major drivers of climate on different temporal-scales.
Trvalý link: https://hdl.handle.net/11104/0341452