Range Specific Neighborhood Rough Set Based Feature Selection for Driver Inattention Classification

Abstract

In the view of performance improvement in machine learning algorithms it is essential to feed them with relevant features. Feature selection is one of the evident process followed by most of the learning algorithms for choosing the relevant features towards reducing the dimensionality of the dataset as well as to improve the classification accuracy. Among various feature selection techniques, Rough Set Theory (RST) has its own major contributions to feature selection domain. However, the conventional rough set based feature selection procedure makes binary decision on either marking an attribute as relevant or irrelevant.The fuzzy based Rough Set could resolve this problem by finding the relevancy by using membership values, however, this method is unable to identify the boundary or range of an attribute value which is appropriate for classification. The idea of feature selection is inappropriate when specific range of an attribute value represents a decision variable while seems to be irrelevant when it’s complete range is considered. This research work focuses on choosing relevant features for the problem of driver inattention detection. The features extracted for the focused problem are real numbers, hence the Neighborhood Rough Set (NRS) model is followed here rather than conventional Rough Set(RS) approach. In this paper, a Range specific Neighborhood Rough Set (RNRS) based feature selection is proposed for more accurate feature selection for the application of detecting driver’s inattention problem. The experiments are carried out with three real time driver datasets and the results are reported to prove the significance of the proposed RNRS based feature selection. Two learning algorithms, namely K-nearest neighbors and support vector machines are usedto evaluate the performance of the proposed approach. Theresults show that the proposed algorithm can significantlyimprove the classification performance.