Research On Multi-label Active Learning Under Weak Labeled Condition

Multi-label classification is a hot topic of machine learning in recent years.Because of the performance of multi-label classification model is highly dependent on the quality and quantity of training examples.Thus,in order to achieve high accuracy of multi-label classification,we need a lot of high-quality labeled examples for training classifier.But as a matter of fact,the process of obtaining labeled examples is usually costly,and especially,the high quality labeled examples are always scarce but unlabeled examples are always massive,therefore active learning can be used to solve such a problem.In this paper,for the particularity of multi-label classification problem under weak labeled condition,we firstly proposed a label dependence exploring method under weak labeled condition,and then proposed two novel active learning frameworks for multi-label classification by exploiting conditional label dependence among labels.We not only take advantage of conditional label dependence for sampling process,but also bring it into constructing a QBC model,to promote active learning process in a semi-supervised way,in order to reduce human annotation cost.We conduct extensive experiments on several real-world datasets and our experimental results show that our new approaches significantly outperforms existing approaches.