B0402
Title: A method for sparse independent component analysis
Authors: Lauri Heinonen - University of Turku (Finland) [presenting]
Joni Virta - University of Turku (Finland)
Abstract: Independent component analysis (ICA) is a popular family of methods for decomposing signals into independent sources. Fourth-order blind identification (FOBI) is an ICA method based on the diagonalization of the kurtosis matrix. A sparse version of FOBI is presented. Compared to regular FOBI, sparse FOBI gives sparse loadings where some of the loadings are estimated to be exactly zero. The FOBI is presented as a sequence of regression problems and the LASSO penalty is used to achieve sparsity. An efficient algorithm for model fitting is given. The method is illustrated with examples and compared to sparse PCA and other relevant competitors.
