Naslov
Do Various Machine Learning Systems Extract the Same Attributes as Relevant Strong Attributes?

Autori
Lukačić, Zoran ; Kern, Josipa ; Gamberger, Dragan

Izvornik
European Notes in Medical Informatics (1861-3179) 1 (2005), 1; 1104-1109

Vrsta, podvrsta i kategorija rada
Radovi u časopisima, članak, znanstveni

Ključne riječi
Artificiel Intelligence; Knowledge Discovery; Machine Learning; Feature Extraction

Sažetak
During the testing reproducibility of relevant attributes sets extracted by two structurally different machine learning systems, See5 and FMLS, a result came out was: between 112 attributes used for description of the real medical problem, both machine learning systems (MLS) detected the same attribute as the most important attribute for problem solving. Systems agreed in the most important attribute, but totally disagreed in the importance of the rest of relevant attributes. It induced the thought that the extraction of some relevant attributes might be independent of the sort of MLS. In this work we added for testing the third system, ILLM, that is structurally different from mentioned two MLS. Once again, the same attribute was extracted by ILLM as the most important for problem solving and once again there was disagreement with other MLS in the rest of relevant attributes. These results show that among attributes used for problem description, might exist two types of attributes relevant for problem solving: (1) attributes (called strong attributes) that can be recognized by MLS as relevant independently of MLS structure and (2) attributes whose recognition as relevant depends of the MLS structure. The consequences of these rules were discussed.

Izvorni jezik
Engleski

Znanstvena područja
Računarstvo, Javno zdravstvo i zdravstvena zaštita, Informacijske i komunikacijske znanosti

POVEZANOST RADA