Interactive improvement of decision trees through flaw analysis and interpretation

Abstract

This paper describes a framework for knowledge acquisition based on analysing and interpreting flaws in decision trees. The decision trees inductively learned are analysed using domain and task specific knowledge to detect improper states called flaws. These are further used to formulate questions to eliminate the flaws by stimulating the acquisition of new examples and domain knowledge for a new induction cycle. To facilitate this process we frame a unified theory in the classification trees' paradigm arguing: (1) what means to have a good/bad tree; (2) why it is good/bad; and (3) how to obtain a better one. We also describe some experimental results of applying this framework to a domain knowledge acquisition system named KAISER and its meta-learner for the decision trees domain theory which build this theory by keeping track of the experts' response of domain level interaction.