Consolidated Trees: An Analysis of Structural Convergence

When different subsamples of the same data set are used to induce classification trees, the structure of the built classifiers is very different. The stability of the structure of the tree is of capital importance in many domains, such as illness diagnosis, fraud detection in different fields, customer’s behaviour analysis (marketing), etc, where comprehensibility of the classifier is necessary. We have developed a methodology for building classification trees from multiple samples where the final classifier is a single decision tree (Consolidated Trees). The paper presents an analysis of the structural stability of our algorithm versus C4.5 algorithm. The classification trees generated with our algorithm, achieve smaller error rates and structurally more steady trees than C4.5 when using resampling techniques. The main focus on this paper is showing how Consolidated Trees built with different sets of subsamples tend to converge to the same tree when the number of used subsamples is increased.