Dimensionality reduction via adjusting data distribution density

Abstract

Dimensionality reduction is an important processing step for pattern recognition. Designing a new optimization goal is a popular method to improve the effect of the dimensionality decrease method. In this paper, we noted that the distribution density of data was not considered in the most classifiers, which may have a negative impact on the classifier. To overcome the above problem, a new optimization goal is designed under the distribution density of the data. In this optimization goal, the sample with smaller density owns larger impact for the optimization result, and then the density of sample could be adjusted to nearly the same in the low dimensional space. The experiments performed verified the proposed method in terms of classification performance.