Title: Proposing a novel multi-label mapping approach for use in SVM-based multi-class classification problems

Authors: Mohammad Saremi; Mehdi Chehel Amirani

Addresses: Faculty of Electrical and Computer Engineering, Urmia University International Campus, Dr. Beheshti Street, Urmia, 5715944514, Iran ' Faculty of Electrical and Computer Engineering, Urmia University, Sero Road, Urmia, 5756151818, Iran

Abstract: In this study, we propose a new method called 'multi-label mapping' (MLM) for solving multi-class classification problems by mapping them into multi-label problems. The MLM method frequently selects different combinations of base classes, merges the classes, and assigns a new label to each class. Six standard datasets are selected from the UCI machine learning repository to evaluate the proposed method. Experimental results demonstrate that the MLM method reduces 50% to 96.66% of the number of required SVM classifiers and 25.76% to 72.27% of the training-testing time in comparison with the OVA and OVO methods. It also yields a better performance in terms of accuracy, precision, recall, and overfitting. Due to the need for a very low number of the SVM binary classifiers, a low training-testing time, and an acceptable prediction error, the presented method is a potential candidate for use in pattern recognition applications and multi-class classification problems.

Keywords: binary classification; binary decision diagram; multi-class classification; multi-label data; pattern recognition; support vector machines; SVM; multi-label mapping; MLM.

DOI: 10.1504/IJDATS.2023.136664

International Journal of Data Analysis Techniques and Strategies, 2023 Vol.15 No.4, pp.255 - 276

Received: 16 Mar 2022
Received in revised form: 27 Feb 2023
Accepted: 28 Jul 2023
Published online: 15 Feb 2024 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article