KIT | KIT-Bibliothek | Impressum | Datenschutz

On a Comprehensive Metadata Framework for Artificial Data in Unsupervised Learning

Dangl, Rainer; Leisch, Friedrich


Evaluating new methods and algorithms in unsupervised learning obviously requires thorough benchmarking studies on data sets that most closely reflect performance in actual usage. Designing data sets that do exactly that is quite a challenging task in itself; standing up to the challenge in comparison to other methods is another point which poses a risk of compromising the goal of an objective benchmarking study. We want to address the latter by proposing a framework that standardizes the format of artificial data, or rather its metadata. We intend to introduce a web repository that functions as an exchange for metadata of artificial data and an accompanying R package that can generate actual data from the descriptions obtained from the repository. It is therefore much simpler to find data designed by others and which has been used in previous benchmarking studies. This removes some of the temptation to specifically design artificial data in a way so that a proposed method performs significantly better than existing ones, a claim that might not hold in real life applications.

Volltext §
DOI: 10.5445/KSP/1000058749/22
Cover der Publikation
Zugehörige Institution(en) am KIT Fakultät für Wirtschaftswissenschaften – Institut für Informationswirtschaft und Marketing (IISM)
Publikationstyp Zeitschriftenaufsatz
Publikationsjahr 2017
Sprache Englisch
Identifikator ISSN: 2363-9881
KITopen-ID: 1000073378
Erschienen in Archives of Data Science, Series A (Online First)
Band 2
Heft 1
Seiten 16 S. online
KIT – Die Forschungsuniversität in der Helmholtz-Gemeinschaft
KITopen Landing Page