Abstract
A new method of feature selection is developed, based on structural similarity. The topological neighborhood information about pairs of objects (or patterns), to partition(s), is taken into consideration while computing a measure of structural similarity. This is termed proximity, and is defined in terms of membership values. Multi-objective evolutionary optimization is employed to arrive at a consensus solution in terms of the contradictory criteria pair involving fuzzy proximity and feature set cardinality. Results for real and synthetic datasets, of low, medium and high dimensionality, show that the method led to a correct selection of the reduced feature subset. Comparative study is also provided, and quantified in terms of accuracy of classification and clustering validity indices.