MPEKDyL: Efficient multi-partial empirical kernel dynamic learning

摘要

Multiple Empirical Kernel Learning (MEKL) is convenient for basic classifiers implemented into feature spaces. However, the huge computational complexity O(MN3) for multiple empirical kernel mapping strongly restricts its application, where N and M are the number of the training samples and the feature spaces, respectively. Moreover, the generated high-dimensional feature spaces result in huge memory cost. To address these problems, we introduce a method called Multi-Partial Empirical Kernel Mapping (MPEKM) that is able to reduce the mapping computational complexity and generate the lower-dimensional feature spaces. The computational complexity OMNl3 could be achieved by this approach, where Nl(≪N) is the size of the partial subset. However, since the feature spaces are constructed by multiple non-overlapping training subsets, they might not represent the discriminant information appropriately. To guarantee the classification performance, we further introduce a Dynamic Learning method (DyL) to dynamically select the samples having more contribution to the decision boundary for training. By combining MPEKM and DyL, we propose an efficient Multi-Partial Empirical Kernel Dynamic Learning method (MPEKDyL) which results in higher classification performance and lower computational complexity than the conventional MEKL. Moreover, the generated feature spaces have much lower dimensions. The advantages of the proposed MPEKDyL are: (i) reducing the multiple empirical kernel mapping computational complexity from O(MN3) to OMNl3, (ii) generating lower-dimensional feature spaces, and (iii) dynamically selecting the samples for training. The experimental results validate its effectiveness and efficiency.