RLS-MARS: An Effective Feature Selection Tool for Text Classification

Abstract

The RLS-MARS (Regularized Least Squares-Multi Angle Regression and Shrinkage) feature selection model is used to select the relevant information, in which both, the keeping and the leaving-out of the regularizer are present. The RLS-MARS model is to find a series of directions in multidimensional space, leading the gradient vectors to change along those directions which would make the gradient matrix's gradient descent, during the procedure, the feature in this direction can be easily selected. TF-IDCFC (Term Frequency Inverse Document and Category Frequency Collection normalization) weighting method is proposed to measure the features, by using category information as a factor. Our experiments on 20Newsgroups and Reuters-21578, all of those results demonstrate the effectiveness of the new feature selection method for text classification.