TY - GEN
T1 - Kernel approaches to unsupervised and supervised machine learning
AU - Kung, Sun Yuan
PY - 2009
Y1 - 2009
N2 - In the kernel approach, any N vectorial or non-vectorial data can be converted to N vectors with feature dimension N. The promise of the kernel approach hinges upon its representation vector space, leading to a "cornerized" data structure. Furthermore, the nonsingular kernel matrix basically assures a theoretically linear separability, critical to supervised learning. The main results are two folds: In terms of unsupervised clustering, the kernel approach allows dimension reduction in the spectral space and, moreover, a simple error analysis for the fast kernel K-means. As to supervised classification, by imposing uncorrelated perturbation to the training vector in the spectral space, a perturbed (Fisher) discriminant analysis (PDA) is proposed. This ultimately leads to a hybrid classier which includes PDA and SVM as specials cases, thus offering more flexibility for improving the prediction performance.
AB - In the kernel approach, any N vectorial or non-vectorial data can be converted to N vectors with feature dimension N. The promise of the kernel approach hinges upon its representation vector space, leading to a "cornerized" data structure. Furthermore, the nonsingular kernel matrix basically assures a theoretically linear separability, critical to supervised learning. The main results are two folds: In terms of unsupervised clustering, the kernel approach allows dimension reduction in the spectral space and, moreover, a simple error analysis for the fast kernel K-means. As to supervised classification, by imposing uncorrelated perturbation to the training vector in the spectral space, a perturbed (Fisher) discriminant analysis (PDA) is proposed. This ultimately leads to a hybrid classier which includes PDA and SVM as specials cases, thus offering more flexibility for improving the prediction performance.
KW - Dimension reduction
KW - Error analysis
KW - Fisher discriminant analysis(FDA)
KW - Kernel matrix
KW - PDA-SVM hybrid classifier
KW - Perturbed discriminant analysis(PDA)
KW - Scatter matrix
KW - Spectral space
KW - Support Vector Machine(SVM)
UR - http://www.scopus.com/inward/record.url?scp=76249132145&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=76249132145&partnerID=8YFLogxK
U2 - 10.1007/978-3-642-10467-1_1
DO - 10.1007/978-3-642-10467-1_1
M3 - Conference contribution
AN - SCOPUS:76249132145
SN - 3642104665
SN - 9783642104664
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 1
EP - 32
BT - Advances in Multimedia Information Processing - PCM 2009 - 10th Pacific Rim Conference on Multimedia, Proceedings
T2 - 10th Pacific Rim Conference on Multimedia, PCM 2009
Y2 - 15 December 2009 through 18 December 2009
ER -