【作者】 肖志博； 王焕钢； 肖英超； 徐文立；

【Author】 Xiao Zhibo Wang Huangang Xiao Yingchao Xu Wenli (Department of Automation,Tsinghua University,Beijing 100084,China)

【机构】 清华大学自动化系；

【摘要】 提出了一种面向大规模数据集的单类支持向量机(OCSVM)方法.该方法基于κ近邻思想得到表征数据集合分布特征的集合内点,并依此生成集合边缘点,而后由二者重新组成数据集合,用于OCSVM训练.该新建数据集不仅极大压缩了原有大规模数据集的样本数量,还可以保留原有大规模数据集的分布特征,从而有效解决了OCSVM在处理大规模数据集时所存在的训练时间长、模型复杂以及预测效率低等问题.最后,通过在典型数据集合上进行的对比实验,表明了所提方法的有效性.更多还原

【Abstract】 A method to train one-class support vector machine(OCSVM)on the large-scale data sets is proposed.The proposed method selects inner points representing the distribution characteristics of the original large-scale data sets based on the principle of k-nearest neighbor,and generates the edge points using the inner points selected.A new data set is formed by combining these two kinds of points to train OCSVM.The new data set not only reduces the volume of the original large-scale data set greatly,but also maintains the distribution characteristics of the original data set.Thus the problems faced by OCSVM on the large-scale data sets,such as long training time,complicated models and low predicting speed,are effectively solved.Finally,the experiment is conducted on typical data sets to illustrate the effectiveness of the method proposed.更多还原