Study on Applicability of Curriculum Framework to Feature Selection Algorithms

Abstract

The present work demonstrates the applicability of consistency based and entropy based feature selection metrics to curriculum framework. Feature selection aims at reducing the dimensionality of datasets and is a preprocessing step for classification and clustering. In general, feature selection algorithms make use of the entire training set for identifying important features. Curriculum based feature selection uses instances in the order of their complexity in an incremental paradigm. Curriculum framework deals with easy and then difficult instances in a guided environment. Curriculum methods are advantageous over no-curriculum methods in two ways. Firstly, the time consumed by curriculum methods is very less; secondly, curriculum methods allow the feature selection methods to be applied in incremental paradigm. Incremental nature of curriculum methods allows feature selection to work upon newly added samples in less time. In this paper, curriculum methods are compared with no-curriculum methods using performance indices - classification accuracy, time and selection ratio. Experimental results on various datasets show (varying degrees of) superiority of using curriculum framework in feature selection.