基于近邻传播聚类的集成特征选择方法  

Affinity Propagation Clustering Based Ensemble Feature Selection Method

作　　者：孟军[1] 尉双云 

机构地区：[1]大连理工大学计算机科学与技术学院,大连116024

出　　处：《计算机科学》2015年第3期241-244,260,共5页Computer Science

基　　金：辽宁省自然科学基金项目(20130200029)资助

摘　　要：针对高维数据中的类标记仅与少部分特征关联紧密的问题,提出了基于排序聚合和聚类分组的特征随机选择集成学习方法。采用排序聚合技术对特征进行过滤,选出与样本分类相关的特征,以bicor关联系数作为关联衡量标准,利用近邻传播聚类算法进行分组,使不同组的特征互不关联,然后从每个分组中随机选择一个特征生成特征子集,便可得到多个既存在差异性又具备区分能力的特征子集,最后分别在对应的特征子空间训练基分类器,采用多数投票进行融合集成。在7个基因表达数据集上的实验结果表明,提出的方法分类误差较低,分类性能稳定,可扩展性好。Aiming at the problem that only a small part of features are associated with the sample classification in highdimensional data containing thousands of features,a filtering and grouping based feature random selection ensemble learning method was proposed.Rank aggregation technique was used to select the relevant features,and we grouped them by affinity propagation clustering algorithm using bicor correlation coefficient as distance measure.The feature clusters were produced and the feature pairs from any two different clusters are not correlated.A feature from each cluster was selected randomly,and then a relevant and discriminative feature subspace was generated.In this way,many feature subspaces can be generated.Base classifiers were trained in the produced feature subspaces and fused together using a majority voting method.The experiments on 7gene expression data sets show that the proposed method can effectively reduce the classification error.Meanwhile,it also has more stable performance,and good expansibility.

关 键 词：分类 排序聚合 近邻传播聚类 集成特征选择 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]