A study on the selection of local training sets for hierarchical classification tasks

In hierarchical classification tasks using the local approach, an important decision concerns the selection of training examples to build the local classifiers. To this end, several policies, which take into account the class taxonomy information, have been proposed. However, a study of a comprehensive comparison concerning the performance of these policies is still lacking. This paper presents a comprehensive empirical evaluation of eight different policies using 13 datasets. The results have shown that three of these policies outperformed the other five policies with statistically significant differences.