Decision tree pruning using an additive information quality measure

An additive quality measure based on information theory is introduced for the inductive inference of decision trees. It takes into account both the information content and the complexity of a tree, combined so as to evaluate the tree on the basis of its learning sample. The additivity of the quality measure with respect to the decomposition of a tree into subtrees, allows to formulate an efficient recursive backward pruning algorithm to maximize the quality. Simulation results are provided on the ground of a real life problem related to electric power system operation and a synthetic digit recognition problem.