一种适于在线学习的增量支持向量数据描述方法

An Incremental Support Vector Data Description Method for Online Learning

本文针对支持向量数据描述(Support Vector Data Description,SVDD)中的在线学习问题,提出了一种增量支持向量数据描述(Incremental Support Vector Data Description,ISVDD)方法。首先,理论明确了增量学习机理在SVDD中的可行性,并深入分析了在线新增样本与已有样本集合的集合划分问题;同时从理论上给出了ISVDD中样本系数变化的依据,推导了ISVDD的理论过程。其次,为了提高理论完备性与应用可靠性,在六种条件下实现了样本属性之间的迁移,获得各个样本系数的变化量。ISVDD方法不仅继承了标准SVDD的优点,能够获得和标准SVDD同样的分类性能,并且显著减少了在线增量样本的训练时间,缓解了数据优化中对内存量的巨大需求。实验结果证明了本文方法的有效性和正确性。

In this paper, an incremental support vector data description （ISVDD） method is proposed for online learn- ing, which is an iterative process for training incremental samples. Firstly, the feasibility of incremental learning on SVDD is proved and we analyze the set division of the incremental sample and existent samples in detail. Meanwhile, the principle of samples＇ coefficient changing is provided and we develop the complete theory of ISVDD. Secondly, in order to improve the theoretical integrality and reliability in application, under six conditions, the migrating of samples can be achieved, which results in the sample coefficients＇ changing. Finally, the whole procedure of ISVDD is proposed to achieve online learning of the ISVDD method. Compared with the standard SVDD, the proposed method not only inherits the excellence of SVDD and can achieve the classification accuracy as same as that of standard SVDD, but also can reduce training time for online incremental samples and release large memory burthen. The experimental results prove the efficiency and validity of our proposed method.