基于凸壳的在线单类学习机

A One-Class Online Classifier Based on Convex Hull

传统的基于支持向量机的单类分类器因计算复杂度高而无法满足大规模数据实时处理的需求,在线学习方法为解决该问题提供了一种有效途径.本文在挖掘样本数据在特征空间分布性状的基础上,提出了一种基于凸壳的在线单类学习机(0ne-class Online Classifier based on Convex Hull,OOCCH).该方法首先使用凸壳的定义选择能代表特征空间中数据分布的凸壳向量对应的原始样本作为训练样本来缩减训练集的规模;其次在分类器在线更新阶段利用凸壳向量动态地调整分类器的训练样本.理论分析证明了OOCCH的有效性,与现有的在线单类分类器的实验比较,OOCCH在训练时间和分类性能方面有显著优势.

Facing the challenge of large-scale data processing,the traditional SVM(support vector machine)based one-class classifier suffers from its high computational complexity.The online learning technique is an effective way to solve this problem.In this paper,a one-class online classifier based on convex hull(OOCCH)is proposed by considering the distribution characteristics of the data in the feature space.In order to reduce the number of training sets,OOCCH selects the samples corresponding to the convex hull vectors in the feature space as training samples.In the online update stage of the classifier,OOCCH dynamically adjusts the training samples based on the definition of convex hull.Theoretical analysis proves the effectiveness of OOCCH.Compared with the existing online one-class classifiers in experiments,OOCCH has significant advantages in training time and classification performance.