PLVI-CE: a multi-label active learning algorithm with simultaneously considering uncertainty and diversity

In multi-label learning, each instance simultaneously associates with multiple labels, which means that labeling such instances is quite costly. Active learning, as an important machine learning paradigm, learns the classification model by querying merely a small portion of data with important information, by means of which, the labeling cost can be greatly reduced during the training process and an accurate and robust classification model could be obtained. Therefore, multi-label active learning (MLAL) has garnered increasing attentions. The primary challenge in MLAL lies in designing an effective query strategy to measure uniform information about unlabeled instances throughout all labels. In this study, we propose a query strategy named predicted label vectors inconsistency and cross entropy measure (PLVI-CE) that considers both uncertainty and diversity measures. In PLVI-CE, the uncertainty is measured by the inconsistency between two predicted label vectors from the same unlabeled instance, and the diversity is assessed by the average discrepancy in posterior probabilities between each unlabeled instance and all instances in the labeled set. Furthermore, in this study, we try to adopt label-weighted extreme learning machine (LW-ELM) as the base classifier in the MLAL framework with considering its following advantages: (1) LW-ELM has a low computational cost, (2) LW-ELM has strong generalization performance, and (3) LW-ELM can be directly used to classify multi-label data with class imbalance distributions, hence providing approximately unbiased instance querying during MLAL. Experimental results on 12 benchmark multi-label datasets indicate the effectiveness and superiority of the proposed PLVI-CE algorithm in comparison with several current state-of-the-art MLAL algorithms.