关键词: 流形学习, 核方法, 多核学习, 原信息, 映射空间

Abstract: It is well known that the manifold structure of data will greatly affect final results about dimensionality reduction. However, most of the existing multiple kernel learning for dimensionality reduction does not consider the problem of maintaining the original manifold structure of data. This paper proposes a new multiple kernel learning for dimensionality reduction method to overcome the possibly distorted manifold structure of data. The core idea of the proposed method is to combine all or partial features of the original data into the multiple kernel expression such that the degree of such distortion is reduced. The combination of original information and feature information will eventually be expressed as a coupling between kernels, and this problem can be optimized through the original multiple kernel learning framework to get its kernel weight coefficients. Experimental results show that the proposed method can achieve a promising result with both a smaller amount of feature descriptor and less running time, in contrast to that of original method this paper based on.

Key words: manifold learning, kernel method, multiple kernel learning, original data, mapping space