CCDH: Complexity based Causal Discovery of Hidden common cause in time series

Abstract

Causal discovery based on temporal observations poses a significant challenge, especially when dealing with causal relationships among dynamical systems exhibiting chaotic attractors. Existing solutions are data-intensive and unable to detect hidden common drivers (also known as common causes or confounders). To address these limitations, we propose a novel method that overcomes both issues. Our method relies on Takens’ embedding theorem and assesses the complexity of rank order patterns in the embedded series by measuring compressibility through non-sequential recursive pair substitution. Remarkably, this approach is effective even with short data samples and has the capability to detect both unidirectional and bidirectional causation, as well as hidden common causes. To validate its performance, we apply the method to synthetic datasets and human electrophysiological data obtained from an epileptic patient. The method successfully provides insights into the involvement of the left and right hippocampus before, during, and after an epileptic seizure. Consequently, our method may offer valuable additional information for decision-making medical panels in determining the optimal intervention locations. Due to its advantages and simplicity, this method holds promise for application in various scientific and practical domains with successful outcomes.