Mapped-PCMCI: an algorithm for causal discovery at the grid level

Abstract

<p>We propose a novel causal discovery method for large-scale gridded time series datasets. Causal discovery has been applied to study a number of problems in climate research in recent years. Causal discovery can be conducted either among spatially aggregated variables (such as modes of climate variability) or by inferring a climate network where the associations among pairs of grid points are treated as a network. In the latter case, causal methods have to deal with several challenges arising from the high dimensionality of such datasets and the data's spatially and temporally redundant nature.</p><p>Our method, called Mapped-PCMCI, aims to overcome some of these challenges. The central idea is based on the assumption that there is a lower-dimensional representation of the causal dependencies among different locations. The method first reconstructs a lower-dimensional spatial representation of the data, then conducts causal discovery utilizing the PCMCI method (Runge. et al. 2019), in that lower-dimensional space, and finally maps causal relations back to the grid level. Using spatiotemporal data generated with the spatially aggregated vector-autoregressive (SAVAR) model (Tibau et al. 2020), we demonstrate that Mapped-PCMCI outperforms state-of-the-art methods in orders of magnitude by utilizing the assumption of a lower-dimensional dependency structure. Mapped-PCMCI can be used to better estimate climate networks and thereby help to understand the climate system from the perspective of complex network theory.</p><p> </p><p>J. Runge, P. Nowack, M. Kretschmer, S. Flaxman, D. Sejdinovic, Detecting and quantifying causal associations in large nonlinear time series datasets. Sci. Adv. 5, eaau4996 (2019).</p><p>Tibau, X.-A., Reimers, C., Eyring, V., Denzler, J., Reichstein, M., and Runge, J.: Spatiotemporal model for benchmarking causal discovery algorithms, EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-9604, https://doi.org/10.5194/egusphere-egu2020-9604, 2020</p>