Artificial linearization in the version space algorithm

Author

Greene, William A.

Author_Institution

Dept. of Comput. Sci., New Orleans Univ., LA, USA

fYear

1991

fDate

7-10 Apr 1991

Firstpage

361

Abstract

A variant of Mitchell´s version space algorithm is studied. The changes to computational cost and predictive accuracy that result from artificially linearizing certain featural dimensions that are in fact nominal (unstructured) are investigated. It is shown that one must be highly selective in choosing which features to artificially linearize. An information theoretic measure is used to identify a small number of features whose value-sets best differentiate between positive and negative training examples. The resulting algorithm´s computational costs approximately double; its predictive accuracy also improves, but, surprisingly, only modestly for the soybean disease dataset which is used as a testbed

Keywords

inference mechanisms; knowledge acquisition; learning systems; Mitchell´s version space algorithm; artificial linearisation; computational cost; inference mechanisms; information theoretic measure; knowledge acquisition; learning systems; negative training examples; positive training examples; predictive accuracy; soybean disease dataset; value-sets; Accuracy; Computational efficiency; Computer science; Diseases; Machine learning; Machine learning algorithms; Shape; Testing;

fLanguage

English

Publisher

ieee

Conference_Titel

Southeastcon '91., IEEE Proceedings of

Conference_Location

Williamsburg, VA

Print_ISBN

0-7803-0033-5

Type

conf

DOI

10.1109/SECON.1991.147773

Filename

147773