CONSOLIDATED TREE CONSTRUCTION ALGORITHM: STRUCTURALLY STEADY TREES
J. M. Pérez, J. Muguerza, O. Arbelaitz, I. Gurrutxaga · 2004
This paper presents a new methodology for building decision trees or classification trees (Consolidated Trees Construction algorithm) that faces up the problem of unsteadiness appearing in the paradigm when small variations in the training set happen. As a consequence, the understanding of the made classification is not lost, making this technique different from techniques such as bagging and boosting where the explanatory feature of the classification disappears. The presented methodology consists on a new metaalgorithm for building structurally more steady and less complex trees (consolidated trees), so that they maintain the explaining capacity and they are faster, but, without losing the discriminating capacity. The meta-algorithm uses C4.5 as base classifier. Besides the meta-algorithm, we propose a measure of the structural diversity used to analyse the stability of the structural component. This measure gives an estimation of the heterogeneity in a set of trees from the structural point of view. The obtained results have been compared with the ones get with C4.5 in some UCI Repository databases and a real application of customer fidelisation from a company of electrical appliances.