On the value of filter feature selection techniques in homogeneous ensembles effort estimation. Issue 6 (30th March 2021)

• Record Type: Journal Article
• Authors: Hosni, Mohamed; Idri, Ali; Abran, Alain
• Is Part Of: Journal of software. Volume 33:Issue 6(2021)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2021-03-30
• Subjects: ensemble effort estimation -- feature selection -- filter -- machine learning -- software development effort estimation; Software engineering -- Periodicals; Computer software -- Development -- Periodicals; Software maintenance -- Periodicals; 005.1
• DOI: 10.1002/smr.2343
• Languages: English
• ISSNs: 2047-7473

Abstract

Abstract: Software development effort estimation (SDEE) remains as the principal activity in software project management planning. Over the past four decades, several methods have been proposed to estimate the effort required to develop a software system, including more recently machine learning (ML) techniques. Because ML performance accuracy depends on the features that feed the ML technique, selecting the appropriate features in the preprocessing data step is important. This paper investigates three filter feature selection techniques to check the predictive capability of four single ML techniques: K ‐nearest neighbor, support vector regression, multilayer perceptron, and decision trees and their homogeneous ensembles over six well‐known datasets. Furthermore, the single and ensembles techniques were optimized using the grid search optimization method. The results suggest that the three filter feature selection techniques investigated improve the reasonability and the accuracy performance of the four single techniques. Moreover, the homogeneous ensembles are statistically more accurate than the single techniques. Finally, adopting a random process (i.e., random subspace method) to select the inputs feature for ML technique is not always effective to generate an accurate homogeneous ensemble. Abstract : Three filter feature selection techniques were investigated to check the predictive capability of four single machine learning techniques and their homogeneous ensembles over six well‐known datasets. The three filters investigated improve the reasonability and the performance of the four single techniques. Moreover, the homogeneous ensembles are statistically more accurate than the single techniques. Finally, adopting a random process to select the inputs feature for machine learning technique is not always effective to generate an accurate homogeneous ensemble.