Learning Categorical Decision Criteria in Biomedical Domains

Abstract

One of the applications of inductive learning has been to derive explicit decision criteria for decision support. This paper reports a machine learning tool called CRLS (Criteria Learning System) which learns decision rules in the form of criteria tables. The mathematical forms of the bias underlying criteria tables are outlined, and the results of testing the tool on real data from several biomedical domains are presented. The success of the experiments is explained in part by the appropriateness of the bias in these domains. This bias is seen to involve two fundamental components unateness and non-equivalence symmetry. These terms are defined and briefly illustrated. Finally, the performance of CRLS is compared with the performance of AQ, contrasting the unate bias with the conjunctive bias in learning categorical decision rules.