An ensemble approach to multi-label classification of textual data

In this paper, we investigate different approaches to multi-label classification of textual data, with a special focus on ensemble techniques. Commonly used classifier ensembles combine outputs of base learning models in order to enhance the learning results. The multi-label classification problem introduces some new challenges to the ensemble learning methods. For instance, one needs to decide in which order is it better to aggregate the base learners - on a level of individual labels and then for the whole label sets, or the other way around. We discuss this issue and experimentally compare selected approaches. In the experiments, we use data from JRS'2012 Data Mining Competition, whose scope was topical classification of biomedical research papers, and as the base learners we utilize the models employed by the winners of this contest.