A New Proposed Feature Subset Selection Algorithm Based on Maximization of Gain Ratio

• Published in: Big Data Analytics pp. 181 - 197
• Main Authors: Nagpal, Arpita, Gaur, Deepti
• Format: Book Chapter
• Language: English
• Published: Cham Springer International Publishing 2015
• Series: Lecture Notes in Computer Science
• Subjects: Classification; Feature selection; Filter method; Mutual information
• ISBN: 9783319270562; 3319270567
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-319-27057-9_13

Feature subset selection is one of the techniques to extract the highly relevant subset of original features from a dataset. In this paper, we have proposed a new algorithm to filter the features from the dataset using a greedy stepwise forward selection technique. The Proposed algorithm uses gain ratio as the greedy evaluation measure. It utilizes multiple feature correlation technique to remove the redundant features from the data set. Experiments that are carried out to evaluate the Proposed algorithm are based on number of features, runtime and classification accuracy of three classifiers namely Naïve Bayes, the Tree based C4.5 and Instant Based IB1. The results have been compared with other two feature selection algorithms, i.e. Fast Correlation-Based Filter Solution (FCBS) and Fast clustering based feature selection algorithm (FAST) over the datasets of different dimensions and domain. A unified metric, which combines all three parameters (number of features, runtime, classification accuracy) together, has also been taken to compare the algorithms. The result shows that our Proposed algorithm has a significant improvement than other feature selection algorithms for large dimensional data while working on a data set of image domain.