Solving multi-class problems with linguistic fuzzy rule based classification systems based on pairwise learning and preference relations

  • Authors:

  • Alberto Fernández;María Calderón;Edurne Barrenechea;Humberto Bustince;Francisco Herrera

  • Affiliations:

  • Department of Computer Science, University of Jaén, Spain;Department of Automatic and Computation, Public University of Navarra, Spain;Department of Automatic and Computation, Public University of Navarra, Spain;Department of Automatic and Computation, Public University of Navarra, Spain;Department of Computer Science and Artificial Intelligence, CITIC-UGR (Research Center on Information and Communications Technology), University of Granada, Spain

  • Venue:
  • Fuzzy Sets and Systems
  • Year:
  • 2010

Quantified Score

Hi-index 0.20

Visualization

Abstract

This paper deals with multi-class classification for linguistic fuzzy rule based classification systems. The idea is to decompose the original data-set into binary classification problems using the pairwise learning approach (confronting all pair of classes), and to obtain an independent fuzzy system for each one of them. Along the inference process, each fuzzy rule based classification system generates an association degree for both of its corresponding classes and these values are encoded into a fuzzy preference relation. Our analysis is focused on the final step that returns the predicted class-label. Specifically, we propose to manage the fuzzy preference relation using a non-dominance criterion on the different alternatives, contrasting the behaviour of this model with both the classical weighted voting scheme and a decision rule that combines the fuzzy relations of preference, conflict and ignorance by means of a voting strategy. Our experimental study is carried out using two different linguistic fuzzy rule learning methods for which we show that the non-dominance criterion is a good alternative in comparison with the previously mentioned aggregation mechanisms. This empirical analysis is supported through the corresponding statistical analysis using non-parametrical tests.