Locally Discriminant Projection with Kernels for Feature Extraction

  • Authors:
  • Jun-Bao Li;Shu-Chuan Chu;Jeng-Shyang Pan

  • Affiliations:
  • Department of Automatic Test and Control, Harbin Institute of Technology, Harbin, China;Department of Information Management, Cheng Shiu University, Kaohsiung, Taiwan;Department of Electronic Engineering, National Kaohsiung University of Applied Sciences, Kaohsiung, Taiwan

  • Venue:
  • ADMA '07 Proceedings of the 3rd international conference on Advanced Data Mining and Applications
  • Year:
  • 2007

Quantified Score

Hi-index 0.00

Visualization

Abstract

Local Preserving Projection (LPP) is an unsupervised feature extraction method which considers the nearest neighbor information and has little to do with the class information, and it fails to perform well for the nonlinear problems due to its limitation of linearity. In this paper, we extend LPP to propose a novel feature extraction namely Kernel Locally Discriminant Projection (KLDP) by considering class label information and the nonlinear problems. The main work lies in: 1) the class label information is considered to create the similarity measure for the local structure graph; 2) the class-wise cosine similarity measure is applied to solve the selection of the free parameter of the similarity measure; 3) kernel method is applied to solve limitation of linearity of LPP. Besides some theory derivation, the experiments are implemented on ORL and Yale face database to evaluate the feasibility of the proposed algorithm.