Feature selection based on feature curve of subclass problem
Lei Liu, Bing Zhang, Shidong Wang, Shuangjie Li, Kaixiang Zhang, Shuqin Wang · 2019
Feature selection is a key step to improve classification performance. Feature selection methods are divided into three types: filters, wrappers and embedded. Generally speaking, the filter methods use one score to judge the comprehensive classification ability of features for all classes. The higher the score is, the stronger the classification ability is. However, studies in many literature have indicated that only by selecting features with high scores often cannot achieve good effect. Therefore, this paper introduces a new feature selection method based on feature curve of subclass problem (Information Gain Regression Curve Feature Selection, referred to as IGRCFS) to find the features with high discriminal ability for each class, and then obtain the optimal feature subset. In order to verify the validity of the IGRCFS method, five kinds of existing feature selection methods are compared on eight datasets. The results demonstrate that the proposed method IGRCFS is effective.