Feature unionization: a novel approach for dimension reduction

Abstract

Dimension reduction is an effective way to improve the classification performance in machine learning. Reducing the irrelevant features decreases the training time and may increase the classification accuracy. Although feature selection as a dimension reduction method can select a reduced feature subset, the size of the subset can be more reduced and its discriminative power can be more improved. In this paper, a novel approach, called feature unionization, is proposed for dimension reduction in classification. Using union operator, this approach combines several features to construct a more informative single feature. To verify the effectiveness of the feature unionization, several experiments were carried out on fourteen publicly available datasets in sentiment classification domain using three typical classifiers. The experimental results showed that the proposed approach worked efficiently and outperformed the feature selection approach.