Abstract:

Similar to traditional learning methods, multilabel learning also suffers from the problems, such as overfitting and the curse of dimensionality, which are raised from high dimensionality of data. Although many multilabel learning algorithms have been proposed, the issue of the high dimensionality has not yet received enough attentions. To solve this problem, we exploit the correlation of features to classify labels by using conditional mutual information, and then perform feature selection on data. Furthermore, a new ensemble learning algorithm for multilabel data is proposed. Experiment results on several multilabel data sets show that the proposed algorithm outperforms the wellestablished multilabel learning algorithms in most cases.

Key words: data mining;multilabel learning;feature selection;conditional mutual information