A novel causality quantification based on information geometry #

Heng Jie Choong, Eun-jin Kim, Fei He

Poster session

Abstract #

Causality quantification is vital in many fields including brain networks, disease spread, etc. The commonly used methods to measure the causality are Granger Causality and Transfer Entropy. Both of these methods are based on measuring the improvement in prediction of one process knowing the other process at an earlier time but have limitations. For instance, Granger Causality relies on linear, stationary processes and parametric modeling at the simplest level while Transfer Entropy requires the calculation of high dimensional probability distribution.

In this talk, we provide a novel, alternative approach based on information geometry [1]. Specifically, based on the concept of the information rate which measures the rate of change of the time-dependent distribution, we further develop the information rate causality as a model-free approach for causality quantification by properly measuring the evolution of the unequal time joint distribution of one process given another process at earlier times. We show that our proposed information causality outperforms Granger causality and Transfer Entropy in linear and nonlinear stochastic systems.

[1] Eun Jin Kim and Adrian Josue Guel-Cortez. “Causal information rate”. In: Entropy 23.8 (2021), pp. 1–20. issn: 10994300. doi: 10.3390/e23081087.