Class-Specific Guided Local Feature Selection for Data Classification

In this paper, a novel class-specific guided local feature selection (CSGLFS) model is proposed, which derives that each area of the high-dimensional data is related to its own unique discriminant feature subset. This algorithm is superior to existing methods which simply applies a common global feature subset to represent all classes in the high-dimensional input data. In the proposed way, the feature subsets take into account local changes, which optimally adjust to the specific region of high-dimensional space. We also give a linked scheme to measure the similarity between query data and each of the specific classes. Moreover, our CSGLFS method is effectively expressed as a linear programming problem. Experimental results on six public datasets prove that our proposed algorithm is capable to obtain meaningful results compared to existing well-known and advanced feature selection algorithms in terms of classification problem.

I. Introduction

High-dimensional data are ubiquitous in various research fields, such as computer vision, bioinformatics and information retrieval. The existence of high-dimensional data increases the demand for computer hardware performance, such as memory and computing power. At the same time, due to the influence of abnormal points or noise, the performance of relevant algorithms is significantly reduced [1], [2]. Thus, numerous dimensionality reduction techniques are applied meaningfully to address these issues by discovering the optimal feature subset to represent original high-dimensional feature vectors. Data processed in this way is beneficial to reduce the running time and make the model to be optimized more compact and general [2].