M. Azam (Austria), F. Berzal (Spain), and K.P. Pfeiffer (Austria)
Classification, ensemble techniques, bagging, trimmedbagging, decision trees, splitting criteria.
Bagging is one of the most successful ensemble classification techniques. Ensembles are used to enhance the predictive capability of unstable classifiers. In fact, it has been proven that bagging outperforms single classifiers [1]. We introduce the idea of selecting those base classifiers in the ensemble whose number of misclassified units is less than or equal to the modal number of misclassified units. This automated mode based bagging technique provides almost the same prediction accuracy than standard bagging, but it does so with a significant reduction in the number of base classifiers in the resulting ensemble. Experimental results using some real-life datasets available from the UCI Machine Learning Repository illustrate the performance of our suggested scheme. Our experiments show that mode-based bagging achieves a significant reduction in the number of base classifiers in the resulting ensembles (44% reduction with respect to standard bagging [1], 25% reduction with respect to trimmed bagging [2]) while the minor increase in the error rates that is observed is not statistically significant. The Friedman test [3] shows that the three studied ensemble methods exhibit a significantly different behavior with respect to the reduction of the number of base classifiers in the ensembles (with p ≤ 0.000041).
Important Links:
Go Back
