Novel Approaches to Identify Clusters Using Independent Components Analysis with Application

Table 1

Main features of the JADE, FastICA, and SOBI.

Category	Method	Algorithms	Pros	Cons
Higher order statistic-based approach	NonGaussianity	JADE	(i) Computationally efficient on the low dimensional datasets in terms of running time requirements	(i) Does not consider the temporal characteristics of the dataset
			(ii) Stable in terms of memory space requirements	(ii) Inefficient for high dimensional datasets in terms of computational speed.
		FastICA	(i) Computationally efficient in terms of running time	(i) Does not consider the temporal characteristics of the dataset
			(ii) Capability for parallel implementation	(ii) Not robust when criteria for nonGaussianity measurement is kurtosis
				(iii) Higher memory space requirements
Second-order statistic-based approach	Temporal dependence	SOBI	(i) Time-delayed covariance matrices of estimated independent components are closest to the diagonal	Does not consider the selection of auto-covariance order
			(ii) Computationally efficient both in terms of memory space and execution time