A comparative study of selection measures on decision tree structures

摘要

This paper is concerned with a comparative investigation of the most commonly used attribute selection measures in the construction of decision trees. We examine the effect of these measures on the resulting tree structures. Our expeirments study these effects against various sampling policies. The emphasis of earlier works in this field has been on the overall size of the tree, pruned or unpruned, in terms of the number of levels and the number of leaf nodes. We take a more informative view of tree structure which takes the functionality of decision trees into consideration. Our structure-evaluating criterion combines classification proportions with the combinatorial structures. We shall demonstrate that the information-based measures outperform the non-information based ones for unpruned trees against calssification proportion thresholds. Among the information-based measures, the information gain appears to be the best. Pruning improves the performance of non-informution based measures we also show thut class of ication performance is not only related to the attribute selection measures but also to the sampling policies.