Shortcomings of transfer entropy and partial transfer entropy: Extending them to escape the curse of dimensionality

Abstract

Transfer entropy (TE) captures the directed relationships between two variables. Partial transfer entropy (PTE) accounts for the presence of all confounding variables of a multivariate system and infers only about direct causality. However, the computation of PTE involves high dimensional distributions and thus may not be robust in case of many variables. In this work, different variants of PTE are introduced, by building a reduced number of confounding variables based on different scenarios in terms of their interrelationships with the driving or response variable. Connectivity-based PTE variants and utilizing the random forests (RF) methodology are evaluated on synthetic time series. The empirical findings indicate the superiority of the suggested variants over TE and PTE, especially in case of high dimensional systems.

arXiv comment13 pages, 14 tables, 2 figures

References in corpus (10)

• Measuring Information Transfer
• Estimating Mutual Information
• Granger causality and transfer entropy are equivalent for Gaussian variables
• On the performance of different synchronization measures in real data: a case study on EEG signals
• Kernel method for nonlinear Granger causality
• Causation Entropy Identifies Indirect Influences, Dominance of Neighbors and Anticipatory Couplings
• Direct coupling information measure from non-uniform embedding
• Graphical modelling of multivariate time series
• Synergy and redundancy in the Granger causal analysis of dynamical networks
• Redundant variables and Granger causality