ENBIS: European Network for Business and Industrial Statistics

Constructing Classification Trees via Data Mining and Design of Experiments Concepts

23 September 2008, 09:00 – 09:20

Abstract

Submitted by

Irad Ben-Gal

Authors

Irad Ben-Gal, Eugene Kagan and Niv Shkolnik

Affiliation

Dept. of Industrial Engineering, Tel-Aviv University

Abstract

We present a generalized approach for the construction of classification trees. The proposed approach is based on both data mining and design-of-experiments concepts.
In particular, we model the different partitions of the target variable and the attributes' set by nodes in a graph and use information theory measures to set the edges' weights. We show that by a proper selection of these measures, the proposed representation generalizes known construction algorithms of classification trees. Moreover, it enables to better balance the tradeoff between construction complexity and classification optimality and often results in shorter trees. Examples will be given.

Return to programme