Rule Extraction from Prediction Models

Abstract

Knowledge Discovery in Databases(KDD) should provide not only predictions but also knowledge such as rules comprehensible to humans. That is, KDD has two requirements, accurate predictions and comprehensible rules. The major KDD techniques are neural networks, statistics, decision trees, and association rules. Prediction models such as neural networks and multiple regression formulas cannot provide comprehensible rules. Linguistic rules such as decision trees and association rules cannot work well when classes are continuous. Therefore, there is no perfect KDD technique. Rule extraction from prediction models is needed for perfect KDD techniques, which satisfy the two KDD requirements, accurate predictions and comprehensible rules. Several researchers have been developing techniques for rule extraction from neural networks. The author also has been developing techniques for rule extraction from prediction models. This paper briefly explains the techniques of rule extraction from prediction models.