.png){kind=icon}
.png){kind=icon}
Return
A maximal accuracy and minimal difference criterion for multiple kernel
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2024.124378.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Base kernel selection, the task of selecting multiple good kernels, is a key issue in multiple kernel learning (MKL) algorithms. This paper introduces a new framework for base strong kernel selection, named multiple random compact Gaussian kernel learning (MRKL), which is efficient and provides good off-the-shelf usability. It employs a random compact Gaussian kernel to generate a number of kernel candidates, which provide complementary information for investigating the data distribution. Then, it selects strong kernels based on a maximal-accuracy-minimal-difference criterion. It addresses the key challenge to MKL algorithms when each base kernel has a single weight and, as a result, different kernels cannot capture the data characteristics in different subspaces of the input space. MRKL is independent of the MKL algorithm, so that it can be combined with any existing MKL algorithm. As examples, we combine MRKL with two popular MKL algorithms, the same weighted-based-kernel combination algorithm and optimization method-based-kernel combination algorithm. We also design a fast active set algorithm for training support vector machine (SVM) with a combined kernel. Empirical results on benchmark datasets demonstrate that MRKL significantly improves existing MKL algorithms and outperforms other state-of-the-art MKL algorithms.
Keywords:
Multiple kernel learning
Random compact Gaussian kernel
Maximal-accuracy-minimal-difference criterion
Active set
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W
Organization
No organization information available