文献类型：期刊文献

英文题名：A novel multiple Nystrom-approximating kernel discriminant analysis

作者：Wang, Zhe[1];Jie, Wenbo[1];Gao, Daqi[1]

机构：[1]E China Univ Sci & Technol, Dept Comp Sci & Engn, Shanghai 200237, Peoples R China

年份：2013

卷号：119

起止页码：385

外文期刊名：NEUROCOMPUTING

收录：;WOS:【SCI-EXPANDED(收录号:WOS:000323851800044)】；

基金：The authors would like to thank Natural Science Foundations of China under Grant Nos. 61272198, 60903091 and 21176077, the Fundamental Research Funds for the Central Universities, and Shanghai Key Laboratory of Intelligent Information Processing of China under Grant No. IIPL-2012-003 for partial support.

语种：英文

外文关键词：Multiple kernel discriminant analysis; Nystrom approximation; Eigenvalue decomposition; Feature extraction; Kernel-based method; Pattern recognition

摘要：Multiple Kernel Discriminant Analysis (MKDA) adopts an ensemble of multiple kernel matrices K(i)s and is supposed to be more flexible and effective than the original Kernel Discriminant Analysis (KDA). However, with n training samples and p kernel matrices K(i)s, MKDA employs pn(2) space units for all the K(i)s in the optimizing process and simultaneously depends on its solving techniques to handle the optimization problem, which would cause a large space and computational complexity and limit the efficiency and applicability. In order to mitigate this problem, this manuscript adopts the Nystrom method approximating K-i and therefore develops a novel Multiple Nystrom-Approximating Kernel Discriminant Analysis (MNKDA). In practice, the proposed MNKDA first adopts m (m << n) samples to generate an approximating kernel matrix (K-i) over tilde for each K-i and forms an ensemble matrix G = Sigma(p)(t=1)mu(t)(K-i) over tilde. Then, MNKDA directly applies the eigenvalue decomposition onto the Nystrom-based ensemble matrix G and reformulates the proposed discriminant analysis as an eigenvalue problem. The experimental results show that the proposed method can achieve an effective and efficient performance than the classical MKDA. The advantages of the proposed MNKDA are (1) expressing the formulation as an eigenvalue problem resolution instead of using commercial softwares; (2) decreasing the space complexity from O(pn(2)) to O(n(2)) and mitigating the computational complexity from O(n(3)) to O(pmn(2)); and (3) providing an alternative multiple kernel learning technique and inheriting the advantage of multiple kernel learning. (C) 2013 Elsevier B.V. All rights reserved.