Active Learning: An Effective Approach of Mining Unlabeled Data
Yanyan Jiang · 2011
Statistical learning technologies are becoming popular and widely applied in most recent years. To better exploit the information of plenty unlabeled data, active learning, which dynamically acquires label from human oracle, is proposed. In this paper, we surveyed the scenario, theory, and algorithms of active learning. Four cate- gories of active learning algorithms are studied: maximizing informativeness, minimizing expected error, mini- mizing version space and their hybrids. Both theory analysis and empirical study substantiate the effectiveness of active learning algorithms, as well as open problems and research insights are presented.