关键词: 分类,排序聚合,近邻传播聚类,集成特征选择

Abstract: Aiming at the problem that only a small part of features are associated with the sample classification in high-dimensional 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 7 gene expression data sets show that the proposed method can effectively reduce the classification error.Meanwhile,it also has more stable performance,and good expansibility.

Key words: Classification,Rank aggregation,Affinity propagation clustering,Ensemble feature selection