Outlier detection in classification based on feature-selection-based regression

An outlier is a datum that is far from other data points in which it occurs. The appearance of outliers results in a complexity to obtain an accurate classification; numerous statistical and machine learning methods have been proposed to identify the outliers. This paper devotes a regression-based algorithm to the detection and identification of outlier before selecting a suitable classifier. The problem is firstly converted to an high-dimensional regression, then a novel method, based on the combination of multiple-correlation-coefficient-based feature selection for dimensional reduction and t-test for sparsification, is proposed, and an iterated algorithm is also given. Performance on simulated numerical data, low-dimensional iris data and high-dimensional DBWorld E-mail data demonstrate the superiority of the proposed method in outlier identification for classification.