Dimensionality Reduction Using Decision-Based Framework for Classification: Sky and Ground

Abstract

In this paper, we address the problem of dimensionality reduction for classification. Classification of data is challenging if its dimension size is high. We propose a decision-based framework for dimensionality reduction using confidence factor as an evaluation measure for generating a relevant feature subset for a specific target. Confidence factor is generated for all features competent for classification using evidence parameters. Evidence parameters are computed based on intersection of classes in the distribution of feature vector and distance between peaks of distribution of feature vector and are combined using Dempster Shafer combination rule. We demonstrate the results of the proposed framework for sky and ground classification using various datasets. The classification in low dimension space is performed retaining the classification accuracy and optimizing computational time.