MTE-MESS: An algorithm for multivariate non-stationary time series causal discovery

TL;DR: In this paper , the authors proposed an information entropy measure -Modified Transfer Entropy (MTE), which can reflect causality more accurately, and added a Quantile-based Causal Significance Test (QCST) and a Conditional Mutual Information Coefficient-based Conditional Independence Test (CMIC-CIT) to the multivariate transfer entropy (MultiTE) algorithm, proposing MTE-MESS algorithm.

Abstract: Multivariate time series causal discovery is a very meaningful but challenging research content. Due to the influence of nonlinearity and non-stationarity of the time series, the performance of traditional algorithms is often unsatisfactory. We propose an information entropy measure -Modified Transfer Entropy (MTE), which can reflect causality more accurately. Further, we take the MTE as the basic measure, and add a Quantile-based Causal Significance Test (QCST) and a Conditional Mutual Information Coefficient-based Conditional Independence Test (CMIC-CIT) to the Multivariate Transfer entropy (MultiTE) algorithm, proposing MTE-MESS algorithm. QCST can remove those insignificant spurious causalities, and CMIC-CIT can identify indirect causality and remove redundancy. Experiments on two time series datasets with known causality and a simulated dataset show that the MTE-MESS algorithm outperforms the baselines in terms of precision and F1-score.