基于高维几何特征的孤立点检测算法  

An outlier detection algorithm based on hyper-geometric feature

作　　者：朱玉全[1] 刘晟[1] 孙金津[1] 

机构地区：[1]江苏大学计算机科学与通信工程学院,江苏镇江212013

出　　处：《江苏大学学报（自然科学版）》2011年第1期84-88,共5页Journal of Jiangsu University：Natural Science Edition

基　　金：国家科技型中小企业技术创新基金资助项目(09C26213203797);江苏省高校自然科学重大基金资助项目(08KJA520001);江苏省"青蓝工程"项目(BK2010331);江苏省六大人才高峰项目(07-E-025)

摘　　要：针对孤立点检测算法不能较好地模拟人工检测过程、未充分考虑待测数据周围样本分布的问题,提出了一种孤立点检测算法DD-SVDD.该算法综合考虑待测样本与目标样本之间的距离,以及待测样本所在区域样本的分布信息,结合距离和平均密度来确定高维特征空间中决策边界附近待测数据的类别.在训练阶段,考虑了决策边界附近目标训练样本的分布,预留训练样本集中边界的部分目标样本并计算其平均密度;在预测阶段,综合使用距离与平均密度对待测样本的归属进行判断.进行了算法的推导,给出了训练阶段、检测阶段的伪代码,并基于UCI机器学习库中的数据进行试验.结果表明,DD-SVDD算法具有有效性,并能达到较高的识别率.To solve the problem that the manual inspection process was badly simulated and the issues were not fully considered in the outlier detection,such as the distribution of the samples surrounding the data to be tested,an outlier detection algorithm called DD-SVDD was proposed.In this algorithm,the distance between the test samples and the target samples and the distribution information of the test sample region were all considered.The distance and the average density were taken into account to determine the type of data tested near the decision-making boundary in high-dimensional feature space.In the training stage,the distribution of the target training samples near the decision-making boundary was considered,and part of the target samples near the concentrated boundary of training sample were set aside,whose average densities were calculated.In the forecasting stage,the attribution of test samples was estimated by using distance and average density synthetically.The algorithmic derivation was carried out,and the codes of training stage and checking stage were given.The experiment based on UCI data was done.The results show that DD-SVDD is effective,and the recognition rate is high.

关 键 词：孤立点检测 支持向量数据描述 半径延伸 平均密度 人工检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]