Dynamic relevance and interdependent feature selection for continuous data

Abstract

Feature Selection (FS) techniques are employed in expert system for selecting relevant and eliminating redundant features. Information-theoretic based FS methods measure non-linear correlation and have gained popularity recently. However, these methods work with discrete data, leading to an additional overhead of discretization while dealing with continuous data. In order to resolve this, the paper proposes a FS method namely, Dynamic Relevance Interdependent Feature Selection (DRIFS) for continuous data which retains the essence of information-theoretic based methods. DRIFS incorporates two major aspects of FS that are ignored by many state of the art methods. First, the method updates the relevance of every candidate features after a new feature is selected and second, the method discriminates between redundant and interdependent features. The performance of DRIFS is compared with six other FS methods, CIFE, CMIM, DISR, ICAP, IGFS and MRMR. Ten real world datasets, which include seven high-dimensional microarray datasets are used for experimental validation. DRIFS performs better in terms of classification accuracy and the number of feature selected. Experimental results also show an improvement in classification accuracy after the application of DRIFS.