%0 Journal Article %T 一种基于椭球体支持向量描述的异常检测方法<br>Weighted hyper-ellipsoidal support vector data description with negative samples for outlier detection %A 姚宇 %A 冯健 %A 张化光 %A 韩克镇< %A br> %A YAO Yu %A FENG Jian %A ZHANG Huaguang %A HAN Kezhen %J 山东大学学报（工学版） %D 2017 %R 10.6040/j.issn.1672-3961.0.2017.180 %X 摘要： 为了解决训练样本数据集中正类、负类样本不平衡的问题,提出一种考虑负类样本信息的加权超椭球体支持向描述方法(weighted hyper-ellipsoidal support vector data description with negative samples, WNESVDD)。 该方法首先引入马氏距离,充分考虑样本分布信息,同时利用正类、负类样本信息建模,融合代价敏感学习思想对不同类样本赋予不同权重。研究结果表明,所提方法可有效减少决策边界包围的空白区域,更好地调整决策边界,而且数据集的利用率明显提高。所提方法应用在University of California at Irvine(UCI)数据集和半导体工业过程数据上的试验结果证明,所提方法具有较强的异常检测能力,相比于同类方法,漏报误报明显减少。<br>Abstract: To solve the influence of the imbalance between positive and negative samples in training sample set, a method named weighted hyper-ellipsoidal support vector data description with negative samples(WNESVDD)was proposed. Mahalanobis distance was introduced such that the information of sample distribution was completely considered. Both normal and negative samples were utilized to modeling. Cost-sensitive learning was introduced to set different weights for different classes. The results showed that the empty areas that decision boundary enclosed were reduced effectively and the decision boundary was refined in the proposed method. The data utilization rate was obviously improved. Several experiments on University of California at Irvine(UCI)data sets and the data set from the semi-conductor manufacturing process were conducted. The experiments results showed that the proposed method had strong ability of anomaly detection, and compared with the similar method, false positives and false negatives were dramatically reduced %K 样本不平衡 %K 马氏距离 %K 超椭球体支持向量描述 %K 边界几何中心 %K 异常检测 %K 空白区域 %K < %K br> %K Mahalanobis distance %K geometric center of boundary %K empty area %K outlier detection %K sample imbalance %K hyper-ellipsoidal support vector support vector data description %U http://gxbwk.njournal.sdu.edu.cn/CN/10.6040/j.issn.1672-3961.0.2017.180