Multi-label Active Learning with Conditional Bernoulli Mixtures

Multi-label learning is an important machine learning task. In multi-label classification tasks, the label space is larger than the traditional single-label classification, and annotations of multi-label instances are typically more time-consuming or expensive to obtain. Thus, it is necessary to take advantage of active learning to solve this problem. In this paper, we present three active learning methods with the conditional Bernoulli mixture ( CBM) model for multi-label classification. The first two methods utilize the least confidence and approximated entropy as the selection criteria to pick the most informative instances, respectively. Particularly, an efficient approximated calculation via dynamic programming is developed to compute the approximated entropy. The third method is based on the cluster information from the CBM, which implicitly takes the advantage of the label correlations. Finally, we demonstrate the effectiveness of the proposed methods through experiments on both synthetic and real-world datasets.