Normalization based dimensionality reduction of symbolic data: A new approach

Many dimensionality reduction schemes have been proposed for conventional data [3,4,7,10], where the feature vectors of the objects are numerical. However, in many pattern recognition problems the feature values are of symbolic type [5,6,9]. A technique has been proposed for the dimensionality reduction of symbolic data [9], in particular for spall data. We propose a simple and an efficient normalization technique for symbolic span data which is built into the dimensionality reduction procedure. The proposed technique transforms the n-d span data to 2-d span data based on Feature Standard Deviation (FSD), Feature Range Value (FRV), Feature Column Grouping (FCG) and Dynamic Feature Sorting (DFS). In pattern recognition, dimensionality reduction is considered an important stage as it helps to reduce the cost of clustering and realizes the same classification results with few features instead of all n-features. Hence, Clustering Tendency Index (CTI) is used to quantify the suitability of the proposed dimensionality reduction procedure. The efficacy of the algorithm is established by experimental studies made on various data sets.