Correlation Based Feature Selection Algorithms for Varying Datasets of Different Dimensionality

Abstract

Curse of dimensionality problem needs to be addressed carefully when designing a classifier. Given a huge dimensional dataset, one interesting problem is the choice of optimal selection of features for classification. Feature selection is an interesting and most optimal solution to the curse of dimensionality problem. Numerous feature selection algorithms have been proposed in the recent past to solve the curse of dimensionality problem but no one stop solution prevails. This paper proposes two novel algorithms for feature selection namely Reverse Piece-wise Correlation Based Feature Selection (RPwCBFS) and Shuffled Piece-wise Correlation Based Feature Selection (SPwCBFS) that divides the feature space into pieces and computes the similarity of feature subsets in reverse order and in random shuffled manner respectively. The proposed algorithms are compared with Fast Correlation Based Feature selection (FCBF), Fast Correlation Based Feature selection # (FCBF#) and Fast Correlation Based Feature selection In Piece (FCBFiP). Standard medium and huge dimensional datasets are used for experimentation purpose. Experimental results prove that the Reverse Piece-wise Correlation Based Feature Selection algorithm (RPwCBFS) and Shuffled Piece-wise Correlation Based Feature Selection algorithm (SPwCBFS) are prominent solution for feature selection when the underlying dataset is medium sized. For huge dimensional datasets, Shuffled Piece-wise Correlation Based Feature Selection algorithm (SPwCBFS) proves to be an optimal choice.