Behavior of Consolidated Trees when using Resampling Techniques

Many machine learning areas use subsampling techniques with different objectives: reducing the size of the training set, equilibrate the class imbalance or non-uniform cost error, etc. Subsampling affects severely to the behavior of classification algorithms. Decision trees induced from different subsamples of the same data set are very different in accuracy and structure. This affects the explanation of the classification; very important in some domains. This paper presents a new methodology for building decision trees. The final classifier is a single decision tree, so that it maintains the explaining capacity of the classification. A comparison in error and structural stability of our algorithm and the C4.5 algorithm is done. The decision trees generated using the new algorithm, achieve smaller error rates and structurally more steady trees than C4.5 when using subsampling techniques.