A case study of strategic induction: the roman numerals data set

摘要

Strategist is an algorithm for decision-tree induction in which attribute selection is based on the evidence-gathering strategies used by doctors. The advantage is that in problem-solving applications of the induced decision tree, the relevance of a selected attribute can be explained in strategic terms. However, as we show in this paper, Strategist's policy of always giving priority to confirming the likeliest outcome class in the current subset of the data set can sometimes lead to the selection of attributes of limited discriminating power. We present a new version of Strategist in which a tactical approach to the selection of a target outcome class reduces its susceptibility to this problem. A new data set for the classification of Roman numbers as correct or incorrect provides a case study of strategic induction in which we examine the algorithm's behaviour with and without this refinement. The new algorithm tends to produce smaller decision trees than its predecessor and is shown to be comparable in accuracy to ID3 on certain data sets.