DocumentCode :
2946622
Title :
From Laplacian Eigenmaps to Kernel Locality Preserving Projections: Equivalence or Improvement?
Author :
Jia, Peng ; Yin, Junsong ; Zhou, Zongtan ; Hu, Dewen
Author_Institution :
Dept. of Autom. Control, Nat. Univ. of Defense Technol., Changsha, China
Volume :
3
fYear :
2009
fDate :
11-12 April 2009
Firstpage :
771
Lastpage :
774
Abstract :
Kernel locality preserving projections (KLPP) and Laplacian eigenmaps (LE) are often taken as two different kinds of approaches in the application of nonlinear dimensionality reduction, but they are more closely related actually than expected. In this paper, KLPP is proved theoretically to solve exactly the same constrained minimization problem as LE. However, the application of KLPP is sensitive to the selections of kernel type and parameters, whereas LE is more efficient and straightforward. Unfolding results on different datasets of the two approaches are presented, together with the comparison of the computation time between KLPP and LE. In our experiments, the actual running time of LE is shorter than that of KLPP, though the time complexity of the two algorithms is comparable. The conclusion of this paper is a beneficial supplement to the nonlinear dimensionality reduction methods system and can be generalized to other algorithms.
Keywords :
computational complexity; eigenvalues and eigenfunctions; learning (artificial intelligence); Laplacian eigenmap; constrained minimization problem; kernel locality preserving projection; manifold learning; nonlinear dimensionality reduction; time complexity; Automation; Constraint theory; Educational institutions; Kernel; Laplace equations; Machine learning; Machine learning algorithms; Manifolds; Mechatronics; Pattern recognition;
fLanguage :
English
Publisher :
ieee
Conference_Titel :
Measuring Technology and Mechatronics Automation, 2009. ICMTMA '09. International Conference on
Conference_Location :
Zhangjiajie, Hunan
Print_ISBN :
978-0-7695-3583-8
Type :
conf
DOI :
10.1109/ICMTMA.2009.210
Filename :
5203315
Link To Document :
بازگشت