Abstract
Dimensionality reduction is one of the key issues of machine learning and data mining, especially for high-dimensional data set. In the literature, there are various dimensionality reduction methods, such as PCA, LDA, and KLDA, and the difference between them mainly lies in the optimization objective. In this paper, we propose a new dimensionality reduction method, whose optimization objective is to maximize the margin between different classes, after projecting the original features into some specific lower-dimensional subspace. The specific subspace is constructed with the help of soft margin support vector machines. Our experiments based on several real-world datasets show that this method improves the performance on classification, and it not only can reduce redundant information in features but also is robust to noise.