Authors:
Subhajit Chakrabarty
1
and
Haim Levkowitz
2
Affiliations:
1
Louisiana State University Shreveport, LA, U.S.A.
;
2
University of Massachusetts Lowell, MA, U.S.A.
Keyword(s):
High Dimension, Independent Component Analysis, Principal Component Analysis, Clustering, Classification, Dimension Reduction, Stability.
Abstract:
Dimensionality reduction of high-dimensional data is often desirable, in particular where data analysis includes visualization – an ever more common scenario nowadays. Principal Component Analysis, and more recently Independent Component Analysis (ICA) are among the most common approaches. ICA may output components that are redundant. Interpretation of such groups of independent components may be achieved through application to tasks such as classification, regression, and visualization. One major problem is that grouping of independent components for high-dimensional time series is difficult. Our objective is to provide a comparative analysis using independent components for given grouping and prediction tasks related to high-dimensional time series. Our contribution is that we have developed a novel semi-supervised procedure for classification. This also provides consistency to the overall ICA result. We have conducted a comparative performance analysis for classification and predi
ction tasks on time series. This research has a broader impact on all kinds of ICA applied in several domains, including bio-medical sensors (such as electroencephalogram), astronomy, financial time series, environment and remote sensing.
(More)