TY - GEN
T1 - Disjoint Subspaces for Common and Distinct Component Analysis
T2 - 53rd Annual Conference on Information Sciences and Systems, CISS 2019
AU - Akhonda, M. A.B.S.
AU - Long, Qunfang
AU - Bhinge, Suchita
AU - Calhoun, Vince D.
AU - Adali, Tulay
N1 - Funding Information:
This work was supported in part by NSF-CCF 1618551, NSF-NCS 1631838, and NIH-NIBIB R01EB005846.
Publisher Copyright:
© 2019 IEEE.
PY - 2019/4/16
Y1 - 2019/4/16
N2 - Data-driven methods based on independent component analysis (ICA) and its extensions, have been attractive for data fusion as they minimize the assumptions placed on the data. Two widely used extensions of ICA, joint ICA (jICA) and multiset canonical correlation analysis prior to joint ICA (MCCA-jICA) fuse data from different datasets by assuming identical mixing matrices. However, these methods typically only take the common features into account within the linked datasets by disregarding the available distinct features, thus limiting their usefulness. In this paper, we propose a new method for fusion based on ICA and canonical correlation analysis (CCA), disjoint subspace framework using ICA (DS-ICA), to identify and extract not only the common but also distinct features, to yield results that better reveal the underlying relationships between the datasets. Separate analyses on the common and distinct subspaces provide flexibility of choosing suitable algorithm and order for each subspace, ensuring more robust results than the competitive methods. We perform and demonstrate performance advantage of DS-ICA through both simulations and application to multiset functional magnetic resonance imaging (fMRI) data collected from healthy controls, as well as patients of schizophrenia performing an auditory odd ball task (AOD).
AB - Data-driven methods based on independent component analysis (ICA) and its extensions, have been attractive for data fusion as they minimize the assumptions placed on the data. Two widely used extensions of ICA, joint ICA (jICA) and multiset canonical correlation analysis prior to joint ICA (MCCA-jICA) fuse data from different datasets by assuming identical mixing matrices. However, these methods typically only take the common features into account within the linked datasets by disregarding the available distinct features, thus limiting their usefulness. In this paper, we propose a new method for fusion based on ICA and canonical correlation analysis (CCA), disjoint subspace framework using ICA (DS-ICA), to identify and extract not only the common but also distinct features, to yield results that better reveal the underlying relationships between the datasets. Separate analyses on the common and distinct subspaces provide flexibility of choosing suitable algorithm and order for each subspace, ensuring more robust results than the competitive methods. We perform and demonstrate performance advantage of DS-ICA through both simulations and application to multiset functional magnetic resonance imaging (fMRI) data collected from healthy controls, as well as patients of schizophrenia performing an auditory odd ball task (AOD).
UR - http://www.scopus.com/inward/record.url?scp=85065156343&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85065156343&partnerID=8YFLogxK
U2 - 10.1109/CISS.2019.8693045
DO - 10.1109/CISS.2019.8693045
M3 - Conference contribution
AN - SCOPUS:85065156343
T3 - 2019 53rd Annual Conference on Information Sciences and Systems, CISS 2019
BT - 2019 53rd Annual Conference on Information Sciences and Systems, CISS 2019
PB - Institute of Electrical and Electronics Engineers Inc.
Y2 - 20 March 2019 through 22 March 2019
ER -