Classifiability based pruning of decision trees

Abstract

Decision tree pruning is useful in improving the generalization performance of decision trees. As opposed to explicit pruning in which nodes are removed from fully constructed decision trees, implicit pruning uses a stopping criteria to label a node as a leaf node when splitting it further would not result in acceptable improvement in performance. The stopping criteria is often also called the pre-pruning criteria and is typically based on the pattern instances available at node (i.e. local information). We propose a new criteria for pre-pruning based on a classifiability measure. The proposed criteria not only considers the number of pattern instances of different classes at a node (node purity) but also the spatial distribution of these instances to estimate the effect of further splitting the node. The algorithm and some experimental results are presented.