面向不均衡数据集的ARSGOS算法  

ARSGOS Algorithm for Imbalanced Data Sets

作　　者：曹鹏[1,2] 栗伟[1,2] 赵大哲[1,2] 

机构地区：[1]东北大学信息科学与工程学院,沈阳110819 [2]东北大学医学影像计算教育部重点实验室,沈阳110179

出　　处：《小型微型计算机系统》2014年第4期818-823,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61001047)资助;中央高校基本科研业务费专项资金项目(N110618001)资助

摘　　要：现实世界中广泛存在着类别分布不均衡的数据,其分类问题已成为数据挖掘领域中的研究热点.提出一种基于高斯混合模型,并结合自适应随机子空间的升采样算法(ARSGOS).该算法使用随机子空间的思想,在不同的子空间下对少数类的样本空间使用高斯混合模型进行参数估计,并根据估计的概率密度函数构造少数类样本数据,从而使新样本数据更加准确真实,避免了噪音数据的生成.该算法通过引入数据子集交叠率,增强数据采样和分类器学习的差异性,提升了在不均衡数据空间下少数类样本的识别能力.通过大量UCI数据集验证,并采用不同的基分类器,与其他常用的重采样算法进行比较.实验结果表明ARSGOS算法有效地提高了少数类的识别率以及整体的分类性能.There widely exists the class imbalanced data in the real world and the classification of the class imbalance become a hot re- search topic in the field of data mining. We propose a method named ARSGOS, which is an over-sampling algorithm based on the distribution of Gaussian mixture model （ GMM ） combined with adaptive random subspace ensemble. The algorithm is based on the i- dea of random subspace, and estimates the distribution of the minority class with GMM in different subspaces. Then the artificial data of minority class is generated according to the probability function estimated, so that the new instances are more accurate and real, a- voiding generating noisy instances. The proposed algorithm increases the diversity in the procedure of the sampling and learning with the overlapping rate of the subsets, and improves the identification ability of the minority class in the imbalance data. We compare ARSGOS with the existing re-sampling methods based on different base classifiers on the UCI data sets. The experimental results show that ARSGOS improves the accuracy of the minority class and the overall classi？ cation performance effectively.

关 键 词：不均衡数据分类 升采样 随机子空间 高斯混合模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]