Classification Based On Local Feature Selection Via Linear Programming
Narges Armanfard, James P. Reilly

This paper presents a novel local feature selection and classification method, which finds the most discriminative features for different regions of the feature space. To this end, we consider each sample of the training set to be a ``representative point" of its associated class. A feature set (possibly different in size and members) is assigned to each representative point. The process of finding a feature set for each representative point is independent of the others and can be performed in parallel. The proposed method makes no assumptions about the underlying structure of the training set; hence the method is insensitive to the distribution of the data over the feature space. The method is formulated as a linear programming optimization problem, which has a very efficient realization. Experimental results demonstrate the viability of the formulation and the effectiveness of the proposed algorithm.