基于多核支持向量数据描述的单类分类方法

One-class Classification Method Based on Multi-kernel Support Vector Data Description

针对异常检测模型中,单核支持向量数据描述存在映射形式单一以及核函数、核参数选择困难的问题,提出一种多核优化组合的支持向量域描述的单类分类方法。在分析多核映射的核空间基础上,建立多核支持向量数据描述模型,以更灵活地描述训练样本在高维特征空间的边界分布情况。采用目标函数的梯度下降法对该模型的多核组合权重进行分步寻优,并引入异常类测试样本来控制和评价分类器的描述精度和推广能力。仿真实验结果表明,该方法具有更好的学习能力和计算效率。

Considering the support vector data descrlpuon information and hard to choose the best kernel and its parameters, the multi-kernel one-class classification with a linear combination of multi-kernel is proposed. The multi-kernel support vector data description model which can descript the data distribution boundary in eigenspace more flexibly is built after analysing the space of multi-kernel mapping. The optimal combination kernels＇ weight is solved by reduced gradient algorithm. Test dataset which includes abnormal samples is introduced to control and evaluate the description accuracy and expansibility of hyper spherical interface. Experimental results show the method has better learning ability and computing efficiency.