基于少量异常数据的最大间隔新奇检测方法

摘要

Traditional novel detection algorithms often only use the normal samples which account for most of the total sample to construct a classifier, the negative class samples are ineffective.To solve this problem, this paper proposed a large margin method that based on a small amount of abnormal data (BSLM).The basic idea was as below: First, a hypersphere should be constructed to contain as many normal instances as possible, and at the same time made sure the margin between the surface of this sphere and normal instances were large enough.In this way, a closed boundary could be attained which surrounded the normal data and was tightly close to the abnormal data.To build such a sphere, only needed to solve a convex optimization problem, which could be efficiently solved through the existing traditional support vector machine (SVM) model with a little change.By some simulation experiments on the datasets of the machine fault detection, medical diagnosis, and the Arabic numeral recognition, the results show that this method can effectively improve the true positive rate, reduce the false positive rate.At the same time, fivefold cross-validation training methods increases the detection stability.%传统的新奇检测算法往往仅利用占样本大多数的正常实例来构造分类器,少量的负类样本基本不能发挥作用.针对此问题,提出一种基于少量负类样本的最大间隔方法,其基本思想是:先构造一个超球面,让它包含尽可能多的正常实例,同时,球表面到正常实例之间的间隔越大越好,从而得到一个围绕正常实例的闭合而又紧贴异常实例的分类边界.建立这样的超球面,只需要解决一个凸的最优化问题,而这个最优化问题可以通过对传统支持向量机模型稍加改变来实现.在机器故障检测、医疗诊断、阿拉伯数字识别等数据集上进行了仿真实验,实验结果表明该方法能够有效地提高检测率,降低误报率;同时五倍交叉验证方法提高了检测的稳定性.