Feature ranking for enhancing boosting-based multi-label text categorization

Boosting algorithms have been proved effective for multi-label learning. As ensemble learning algorithms, boosting algorithms build classifiers by composing a set of weak hypotheses. The high computational cost of boosting algorithms in learning from large volumes of data such as text categorization datasets is a real challenge. Most boosting algorithms, such as AdaBoost.MH, iteratively examine all training features to generate the weak hypotheses, which increases the learning time. RFBoost was introduced to manage this problem based on a rank-and-filter strategy in which it first ranks the training features and then, in each learning iteration, filters and uses only a subset of the highest-ranked features to construct the weak hypotheses. This step ensures accelerated learning time for RFBoost compared to AdaBoost.MH, as the weak hypotheses produced in each iteration are reduced to a very small number. As feature ranking is the core idea of RFBoost, this paper presents and investigates seven feature ranking methods (information gain, chi-square, GSS-coefficient, mutual information, odds ratio, F1 score, and accuracy) in order to improve RFBoost's performance. Moreover, an accelerated version of RFBoost, called RFBoost1, is also introduced. Rather than filtering a subset of the highest-ranked features, FBoost1 selects only one feature, based on its weight, to build a new weak hypothesis. Experimental results on four benchmark datasets for multi-label text categorization) Reuters-21578, 20-Newsgroups, OHSUMED, and TMC2007(demonstrate that among the methods evaluated for feature ranking, mutual information yields the best performance for RFBoost. In addition, the results prove that RFBoost statistically outperforms both RFBoost1 and AdaBoost.MH on all datasets. Finally, RFBoost1 proved more efficient than AdaBoost.MH, making it a better alternative for addressing classification problems in real-life applications and expert systems.