Simplifying classification trees through consensus methods

Abstract

Methods for comparing and combining classification trees based on proximity measures have been proposed in the last few years. These methods could be used to analyse a set of trees obtained from independent data sets or from resampling methods like bootstrap or cross validation applied to the same training sample. In this paper we consider, as an alternative to the pruning techniques, a modified version of a consensus algorithm we have previously proposed that combines trees obtained by bootstrap samples. This consensus algorithm is based on a dissimilarity measure recently proposed. Experimental results are provided to illustrate, in two real data sets, the performances of the proposed consensus method.