Learning positive-negative rule-based fuzzy associative classifiers with a good trade-off between complexity and accuracy
MetadataShow full item record
AuthorBiedma Rodríguez, Carmen; Gacto, María José; Anguita-Ruiz, Augusto; Alcalá Fernández, Rafael; Aguilera García, Concepción María; Alcalá Fernández, Jesús
Fuzzy associative classificationEvolutionary fuzzy systemseXplainable artificial intelligenceTransparencyComplexity
C. Biedma-Rdguez, M.J. Gacto, A. Anguita-Ruiz, R. Alcalá, C.M. Aguilera, J. Alcalá-Fdez. Learning positive-negative rule-based fuzzy associative classifiers with a good trade-off between complexity and accuracy. Fuzzy Sets and Systems 465 (2023) 108511
SponsorshipFunding for open access charge: Universidad de Granada / CBUA.; ERDF and the Regional Government of Andalusia/Ministry of Economic Transformation, Industry, Knowledge and Universities (grant numbers P18-RT-2248 and B-CTS-536-UGR20); ERDF and Health Institute Carlos III/Spanish Ministry of Science, Innovation and Universities (grant number PI20/00711); Spanish Ministry of Science and Innovation (grant number PID2019-107793GB-I00)
Nowadays, the call for transparency in Artificial Intelligence models is growing due to the need to understand how decisions derived from the methods are made when they ultimately affect human life and health. Fuzzy Rule-Based Classification Systems have been used successfully as they are models that are easily understood by models themselves. However, complex search spaces hinder the learning process, and in most cases, lead to problems of complexity (coverage and specificity). This problem directly affects the intention to use them to enable the user to analyze and understand the model. Because of this, we propose a fuzzy associative classification method to learn classifiers with an improved trade-off between accuracy and complexity. This method learns the most appropriate granularity of each variable to generate a set of simple fuzzy association rules with a reduced number of associations that consider positive and negative dependencies to be able to classify an instance depending on the presence or absence of certain items. The proposal also chooses the most interesting rules based on several interesting measures and finally performs a genetic rule selection and adjustment to reach the most suitable context of the selected rule set. The quality of our proposal has been analyzed using 23 real-world datasets, comparing them with other proposals by applying statistical analysis. Moreover, the study carried out on a real biomedical research problem of childhood obesity shows the improved trade-off between the accuracy and complexity of the models generated by our proposal.