Studies on Reducing the Necessary Data Size for Rule Induction from the Decision Table by STRIM

Abstract

STRIM (Statistical Test Rule Induction Method) has been proposed for an if-then rule induction method from the decision table. STRIM judges the significance of a trying rule by a statistical test based on the table. The method judging the trying rule has been executed based on the standard normal distribution approximating the distribution of the decision attribute’s values so that the judging method needs the proper size dataset satisfying the conditions of the approximation. This paper proposes a new STRIM named minor-STRIM not incorporating the test by the approximating distribution but by the original distribution, which expands the applicable range to cases not satisfying the conditions. Specifically, minor-STRIM uses a binomial distribution for the testing and shows the applicable range expanded and performance evaluation by use of a simulation experiment compared with those by the conventional STRIM. The simulation also shows that it gives discussing and confirming information the validity of the results obtained from applying minor-STRIM to a real-world dataset.