Hierarchical Independent Component Analysis: a multi-resolution non-orthogonal data-driven basis

Keywords

Statistical learning
Code:
01/2014
Title:
Hierarchical Independent Component Analysis: a multi-resolution non-orthogonal data-driven basis
Date:
Monday 13th January 2014
Author(s):
Secchi, P.; Vantini, S.; Zanini, P.
Download link:
Abstract:
We introduce a new method, named HICA (Hierarchical Independent Component Analysis), suited to the dimensional reduction and the multi-resolution analysis of high dimensional and complex data. HICA solves a Blind Source Separation problem by integrating Treelets with Independent Component Analysis and provides a multi-scale non-orthogonal data-driven basis apt to meaningful data representations in reduced spaces. We describe some theoretical properties of HICA and we test the method on synthetic data. Finally, we apply HICA to the analysis of EEG traces.