最大化ROC曲线下面积的不平衡基因数据集差异表达基因选择算法

发布时间：2018-03-05 19:16

本文选题：基因选择　切入点：差异表达基因　出处：《陕西师范大学学报(自然科学版)》2017年01期 　论文类型：期刊论文

【摘要】：针对ARCO(AUC and rank correlation coefficient optimization)算法在进行两类问题特征选择时,采用斯皮尔曼等级相关系数度量已选特征子集冗余性带来信息损失和特征相关性与冗余性度量取值范围不一致的缺陷,提出改进的Pearson相关系数度量特征冗余性,并归一化特征相关性和冗余性度量范围,得到APCO(AUC and improved Pearson correlation coefficient optimization)算法以克服ARCO算法的不足。同时,针对实现多类特征选择的MAUCD(using MAUC as the relevance metric to rank features directly)和MDFS(MAUC decomposition based feature selection method)算法没有考虑特征冗余,且MDFS易选择到局部最优特征子集的问题,提出适于多类问题的改进Pearson相关系数度量特征冗余性,得到基于mRMR(maximal relevance-minimal redundancy)框架的MAUCP和MDFSP算法,克服MAUCD和MDFS算法的缺陷。以SVM、NB和KNN为分类工具,构造基于所选特征子集的相应分类器,以其AUC(MAUC)值度量相应特征子集的性能。7个二类和3个多类不平衡基因数据集的实验结果表明:提出的APCO、MAUCP和MDFSP算法分别优于ARCO、MAUCD和MDFS算法,也优于其他经典基因选择算法。
[Abstract]:Aiming at the defects of the ARCO(AUC and rank correlation coefficient optimization algorithm, which uses the Spelman rank correlation coefficient to measure the redundancy of the selected feature subset, the information loss and the inconsistent range of the feature correlation and redundancy measures are obtained when the ARCO(AUC and rank correlation coefficient optimization algorithm is used to select the two kinds of features. This paper proposes an improved Pearson correlation coefficient to measure feature redundancy, and normalizes the measurement range of feature correlation and redundancy, and obtains the APCO(AUC and improved Pearson correlation coefficient optimization algorithm to overcome the shortcomings of ARCO algorithm. To solve the problem that MAUCD(using MAUC as the relevance metric to rank features directly) and MDFS(MAUC decomposition based feature selection method realize multi-class feature selection without considering feature redundancy, MDFS is easy to select to the locally optimal feature subset. An improved Pearson correlation coefficient for multi-class problems is proposed to measure feature redundancy. The MAUCP and MDFSP algorithms based on the framework of mRMR(maximal relevance-minimal are proposed to overcome the shortcomings of the MAUCD and MDFS algorithms. SVMNB and KNN are used as classification tools. A corresponding classifier based on the selected feature subset is constructed, and the performance of the corresponding feature subset is measured by its AUCMAUC value. The experimental results of 7 classes and 3 classes of unbalanced gene data sets show that the proposed algorithms are superior to ARCOOMAUCD and MDFS algorithms, respectively. It is also superior to other classical gene selection algorithms.
【作者单位】：陕西师范大学计算机科学学院;
【基金】：陕西省科技攻关项目(2013K12-03-24) 国家自然科学基金(61673251) 中央高校基本科研业务费专项资金(GK201503067)
【分类号】：Q811.4

【相似文献】