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

Area of Science:

• Computational neuroscience
• Information dynamics
• Network analysis

Background:

• Information dynamics quantifies statistical structures in complex networks.
• Granger Causality (GC) is a key measure, often computed using Vector Autoregressive (VAR) and state space (SS) models identified by Ordinary Least Squares (OLS).
• Existing methods struggle with limited data samples.

Purpose of the Study:

• To propose a novel identification approach for VAR and SS models using Least Absolute Shrinkage and Selection Operator (LASSO).
• To evaluate LASSO's performance against OLS, particularly in data-scarce conditions.
• To demonstrate the utility of LASSO for computing information dynamics measures.

Main Methods:

• Utilized Least Absolute Shrinkage and Selection Operator (LASSO) for VAR and SS model identification.
• Compared LASSO with Ordinary Least Squares (OLS) through simulation studies.
• Validated the approach on real electroencephalographic (EEG) signals during a motor imagery task.

Main Results:

• LASSO identification maintains accuracy with limited data samples.
• LASSO yields sparse matrices of estimated information transfer.
• LASSO demonstrated superior performance in network structure identification under data paucity compared to OLS.
• LASSO-based analysis of EEG data showed improved network structure.

Conclusions:

• LASSO provides a robust alternative for identifying VAR and SS models in information dynamics.
• This method enhances the analysis of complex network temporal dynamics, especially with limited data.
• The study paves the way for broader application of LASSO regression in computational neuroscience for information dynamics measures.