Unsupervised approach for structure preserving dimensionality reduction

In this paper, a new technique is presented that reduces the dimensionality of large data set without disturbing its topology and also maintains high accuracy of classification. For this Genetic Algorithm is used with Sammon error as the fitness function. The proposed technique is tested on four real and one synthetic data set. High value of correlation coefficient between proximity matrix of original data set and the corresponding data set with reduced number of features ensures that topology of the data set is preserved even with reduced number of features. Comparative study of the clustering results obtained with reduced and original data set justifies the capability of the proposed technique to give good classification accuracy even with reduced number of features.