New boosting algorithms for text categorization

AdaBoost.SZ is a boosting method specifically designed for solving multi-class, multi-label text categorization problems. Fabrizio Sebastiani et al. (2000) provided another idea to improve these base classifiers: combining two or more weak hypotheses as a single base classifier. Its main problem is that the amount of hypotheses selected to combine is determined not by their importance, but by the boosting iteration times already performed. This paper proposes two dynamical ways for combining any number of hypotheses according to their importance. Experimental results show that the new ideas do improve the performance of boosting.