A Bayesian Imprecise Classification method that weights instances using the error costs

Moral-García, Serafín; Coolen-Maturi, Tahani; Coolen, Frank P.A.; Abellán, Joaquín

Abstract

In practical applications, Bayesian classification methods have been successfully employed. The Naïve Bayes algorithm (NB) is a quick, successful, and well-known Bayesian classification method. The Naïve Credal Classifier (NCC) is a version of NB that outputs imprecise predictions (sets of class values). NCC was also adapted for considering classification error costs. Such an adaptation is the only Bayesian method for Imprecise Classification proposed so far that considers misclassification costs. This paper presents a Bayesian algorithm for Imprecise Classification that weights the instances using the misclassification costs in such a way that the importance of an instance increases as the error cost of its class value is higher. We highlight that our proposal may provide more informative and intuitive outcomes than the existing cost-sensitive NCC. We experimentally show that our new proposed method improves the existing cost-sensitive NCC. Moreover, we highlight that our imprecise classifier has a processing time equivalent to the original NB algorithm for precise classification, which has been successfully applied to very large and real datasets. This is a crucial point in favor of our proposal because of the huge amount of data in many application areas nowadays.

Citation

Moral-García, S., Coolen-Maturi, T., Coolen, F. P., & Abellán, J. (2024). A Bayesian Imprecise Classification method that weights instances using the error costs. Applied Soft Computing, 165, 112080. https://doi.org/10.1016/j.asoc.2024.112080

• Journal Article Type: Article
• Acceptance Date: Jul 31, 2024
• Online Publication Date: Aug 13, 2024
• Publication Date: 2024-11
• Deposit Date: Aug 23, 2024
• Publicly Available Date: Aug 23, 2024
• Journal: Applied Soft Computing
• Print ISSN: 1568-4946
• Publisher: Elsevier
• Peer Reviewed: Peer Reviewed
• Volume: 165
• Pages: 112080
• DOI: https://doi.org/10.1016/j.asoc.2024.112080
• Public URL: https://durham-repository.worktribe.com/output/2757087