A comparative study of some issues concerning algorithm recommendation using ranking methods

• Type: Article in International Scientific Journal
• Date: 2002

Authors

Soares, C (Author) FEP; Brazdil, P (Author) FEP

Journal

Title: Lecture Notes in Computer Science

Vol. 2527

Pages: 80-89

ISSN: 0302-9743

Publisher: Springer Nature

Indexing

ISI Web of Knowledge - 4 Citations

Scopus - 4 Citations

Scientific classification

FOS: Natural sciences > Computer and information sciences

Other information

Authenticus ID: P-000-QSJ

DOI: 10.1007/3-540-36131-6_9

Abstract (EN): Cross-validation (CV) is the most accurate method available for algorithm recommendation but it is rather slow. We show that information about the past performance of algorithms can be used for the same purpose with small loss in accuracy and significant savings in experimentation time. We use a meta-learning framework that combines a simple IBL algorithm with a ranking method. We show that results improve significantly by using a set of selected measures that represent data characteristics that permit to predict algorithm performance. Our results also indicate that the choice of ranking method as a smaller effect on the quality of recommendations. Finally, we present situations that illustrate the advantage of providing recommendation as a ranking of the candidate algorithms, rather than as the single algorithm which is expected to perform best.

Language: English

Type (Professor's evaluation): Scientific

No. of pages: 10