@unpublished{eprints_etsmtl3378,
            year = {2023},
           title = {Dynamic ensemble selection using fuzzy min-max hyperboxes},
            note = {"Manuscript-based thesis presented to {\'E}cole de technologie sup{\'e}rieure in partial fulfillment for a master?s degree with thesis in software engineering".
Comprend des r{\'e}f{\'e}rences bibliographiques (pages 139-148).},
       publisher = {{\'E}cole de technologie sup{\'e}rieure},
           month = {novembre},
          author = {Reza Davtalab},
          school = {{\'E}cole de technologie sup{\'e}rieure},
        keywords = {ensemble de classifieurs, s{\'e}lection d?ensemble dynamique, hyperbo{\^i}tes floues, {\'e}chantillons mal class{\'e}s, comp{\'e}tence du classifieur},
        abstract = {Dynamic Selection systems are a good alternative to achieve higher accuracy in complex problems. In current DS techniques, the competence of base classifiers to classify the new query sample is estimated with regard to their performance in a small region surrounding the query sample. Most of these techniques use the KNN algorithm to estimate competence. However, the KNN algorithm endures a high complexity to the system and is also sensitive to the local data distribution. 

In this project, we are going to introduce a novel Dynamic Ensemble Selection framework that uses Fuzzy Hyperboxes to estimate the competence of base classifiers. For the construction of hyperboxes, the distribution of samples is considered and several samples that are close to each other are represented by a hyperbox. Therefore, the final system will not be sensitive to the local imbalance distribution of samples. Besides, the system needs only to keep the hyperboxes rather than the original data. Furthermore, hyperboxes are made based on misclassified samples. Thus, they are usually much fewer than samples. So we expect the proposed framework to have much lower complexity and higher accuracy compared to KNN-based techniques. Experimental results show that the suggested framework can improve the performance of DS systems in both terms of accuracy and computational complexity.},
             url = {https://espace.etsmtl.ca/id/eprint/3378/}
}