Reducing rankings of classifiers by eliminating redundant classifiers

• Type: Article in International Conference Proceedings Book
• Date: 2001

Authors

Brazdil, P (Author) FEP; Soares, C (Author) FEP; Pereira, R (Author) Other

Conference proceedings International

Title: Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

Pages: 14-21

10th Portuguese Conference on Artificial Intelligence, EPIA 2001

Porto, 17 December 2001 through 20 December 2001

Indexing

ISI Web of Knowledge

Scopus - 5 Citations

Other information

Authenticus ID: P-008-6VS

DOI: 10.1007/3-540-45329-6_5

Abstract (EN): Several methods have been proposed to generate rankings of supervised classification algorithms based on their previous performance on other datasets [8,4]. Like any other prediction method, ranking methods will sometimes err, for instance, they may not rank the best algorithm in the first position. Often the user is willing to try more than one algorithm to increase the possibility of identifying the best one. The information provided in the ranking methods mentioned is not quite adequate for this purpose. That is, they do not identify those algorithms in the ranking that have reasonable possibility of performing best. In this paper, we describe a method for that purpose. We compare our method to the strategy of executing all algorithms and to a very simple reduction method, consisting of running the top three algorithms. In all this work we take time as well as accuracy into account. As expected, our method performs better than the simple reduction method and shows a more stable behavior than running all algorithms. © Springer-Verlag Berlin Heidelberg 2001.

Language: English

Type (Professor's evaluation): Scientific