Study On Improved Feature Selection Algorithm Based On Effective Range

In many applications such as pattern recognition, computer vision and data mining, thedata usually has a high dimension, however, these high-dimensional data often containabundant of irrelevant or redundant features, these features will deteriorate the performance ofthe machine learning algorithms, in order to solve this problem, the feature selectiontechniques have attracted many attentions in the past decades. Feature selection can achievethe purpose of reducing the dimensions of the data by removing irrelevant or redundantfeatures. The results of feature selection can directly affect the classifier’s classificationaccuracy and generalization performance.Recently, B. Chandra et al. proposed a statistical feature selection method namedEffective Range based Gene Selection (ERGS). The basic principle behind ERGS is thathigher weight is given to the feature that discriminates the classes clearly. Experimentalresults on several gene expression datasets illustrate the effectiveness of ERGS. However,ERGS method only takes into account the relationship among the effective range of eachfeature for every class, that is overlapping area (OA)，using it to measure whether importantor not a feature is, but neglecting another relationship can also have an impact on the feature’sweight, namely inclusion relation. In order to solve the problems above, a novel efficientstatistical feature selection approach called Improved Feature Selection based on EffectiveRange (IFSER) is proposed in this paper. In IFSER, an including area (IA) is introduced tocharacterize the inclusion relation of effective ranges. Moreover, the samples’ proportion foreach feature of every class in both OA and IA is also taken into consideration. By consideringthe overlapping area (OA), including area (IA) and samples’ proportion to evaluate theimportance of features, thus more conducive to choose the optimal feature subset.Therefore, IFSER outperforms the original ERGS and some other state of the artalgorithms. Extensive experiment results show the effectiveness of the method in this paper.