Dimensionality Reduction may result in contradicting effects- the advantage of minimizing the number of features coupled with the disadvantage of information loss leading to incorrect classification or clustering. This could be the problem when one tries to extract all classes present in a high dimensional population. However in real life, it is often observed that one would not be interested in looking into all classes present in a high dimensional space but one would focus on one or two or few classes at any given instant depending upon the purpose for which data is analyzed. The proposal in this research work is to make the dimensionality reduction more effective, whenever one is interested specifically in a target class, not only in terms of minimizing the number of features, also in terms of enhancing the accuracy of classification particularly with reference to the target class. The objective of this research work hence is to realize effective feature subsetting supervised by the specified target class. A multistage algorithm is proposed- in the first stage least desired features which do not contribute substantial information to extract the target class are eliminated, in the second stage redundant features are identified and are removed to overcome redundancy, and in the final stage more optimal set of features are derived from the resultant subset of features. Suitable computational procedures are devised and reduced feature sets at different stages are subjected for validation. Performance is analysed through extensive experiments. The multistage procedure is also tested on a hyperspectral AVIRIS Indiana Pine data set.

Feature sub setting Target Class Sum of Squared Error Stability Factor Convergence Inference Factor and Homogeneity.