Information Dynamics Analysis: A new approach based on Sparse Identification of Linear Parametric Models

Abstract

The framework of information dynamics allows to quantify different aspects of the statistical structure of multivariate processes reflecting the temporal dynamics of a complex network. The information transfer from one process to another can be quantified through Transfer Entropy, and under the assumption of joint Gaussian variables it is strictly related to the concept of Granger Causality (GC). According to the most recent developments in the field, the computation of GC entails representing the processes through a Vector Autoregressive (VAR) model and a state space (SS) model typically identified by means of the Ordinary Least Squares (OLS). In this work, we propose a new identification approach for the VAR and SS models, based on Least Absolute Shrinkage and Selection Operator (LASSO), that has the advantages of maintaining good accuracy even when few data samples are available and yielding as output a sparse matrix of estimated information transfer. The performances of LASSO identification were first tested and compared to those of OLS by a simulation study and then validated on real electroencephalographic (EEG) signals recorded during a motor imagery task. Both studies indicated that LASSO, under conditions of data paucity, provides better performances in terms of network structure. Given the general nature of the model, this work opens the way to the use of LASSO regression for the computation of several measures of information dynamics currently in use in computational neuroscience.