CFE 2019: Start Registration
View Submission - CMStatistics
Title: Blind source separation of graph signals Authors:  Jari Miettinen - Aalto University (Finland) [presenting]
Sergiy Vorobyov - Aalto University (Finland)
Esa Ollila - Aalto University (Finland)
Abstract: The blind source separation (BSS) problem can be solved either by using non-Gaussianity of the latent components or dependence between the samples. The dependence can be structural such as in time series, spatial data or tensor data, but the focus is on BSS of graph signals which may have more complicated dependencies characterized by graphs and their adjacency matrices. So far, only one BSS method, called GraDe (graph decorrelation), has been designed for this setup. It uses joint approximate diagonalization of graph autocovariance matrices, which are generalizations of autocovariance matrices for time series, and thus GraDe can be seen as generalization of SOBI (second-order blind identification) method for BSS of time series. Modifications of GraDe are suggested, and combining it to non-Gaussianity based methods, FastICA and JADE (joint approximate diagonalization of eigenmatrices), is proposed. In a simulation study, the proposed methods are shown to always achieve at least the performance of the better component (GraDe or FastICA/JADE), and in the case of non-Gaussian graph signals to outperform them both.