Abstract:

In comparison with single kernel learning， Multiple Kernel Learning （MKL） methods obtain better performance in the tasks of classification and regression. However， all the traditional MKL methods are used for tackling two-class or multi-class classification problems. To make MKL methods fit for dealing with the problems of One-Class Classification （OCC）， a Centered Kernel Alignment （CKA） based multiple kernel One-Class Support Vector Machine （OCSVM） was proposed. Firstly，CKA was utilized to calculate the weight of each kernel matrix， and the obtained weights were used as the linear combination coefficients to linearly combine different types of kernel functions to construct the combination kernel function and introduce them into the traditional OCSVM to replace the single kernel function. The proposed method can not only avoid the selection of kernel function， but also improve the generalization and anti-noise performances. In comparison with other five related methods including OCSVM，Localized Multiple Kernel OCSVM （LMKOCSVM） and Kernel-Target Alignment based Multiple Kernel OCSVM （KTA-MKOCSVM） on 20 UCI benchmark datasets， the geometric mean （g-mean） values of the proposed algorithm were higher than those of the comparison methods on 13 datasets. At the time， the traditional single kernel OCSVM obtained better results on 2 datasets，LMKOCSVM and KTA-MKOCSVM achieved better classification effects on 5 datasets. Therefore， the effectiveness of the proposed method was sufficiently verified by experimental comparisons.

Key words: Multiple Kernel Learning (MKL), Centered Kernel Alignment (CKA), One-Class Support Vector Machine (OCSVM), One-Class Classification (OCC), kernel function

摘要：

多核学习（MKL）方法在分类及回归任务中均取得了优于单核学习方法的性能，但传统的MKL方法均用于处理两类或多类分类问题。为了使MKL方法适用于处理单类分类（OCC）问题，提出了基于中心核对齐（CKA）的单类支持向量机（OCSVM）。首先利用CKA计算每个核矩阵的权重，然后将所得权重用作线性组合系数，进而将不同类型的核函数加以线性组合以构造组合核函数，最后将组合核函数引入到传统OCSVM中代替单个核函数。该方法既能避免核函数的选取问题，又能提高泛化性能和抗噪声能力。在20个UCI基准数据集上与其他五种相关方法进行了实验比较，结果表明该方法在13个数据集上的几何均值（g-mean）均高于其他对比方法，而传统的单核OCSVM仅在2个数据集上的效果较好，局部多核单类支持向量机（LMKOCSVM）和基于核目标对齐的多核单类支持向量机（KTA-MKOCSVM）在5个数据集上的分类效果较好。因此，通过实验比较充分验证了所提方法的有效性。

关键词: 多核学习, 中心核对齐, 单类支持向量机, 单类分类, 核函数