Advances in Active Learning Algorithms Based on Sampling Strategy

The classifier in active learning algorithms is trained by choosing the most informative unlabeled instances for human experts to label. In the cycling procedure, the classification accuracy of the model is improved, and then the classifier with high generalization capability is obtained by minimizing the totally labeling cost. Active learning has attracted attentions of researchers both at home and abroad widely. It is pointed out that the active learning technique is a very important research at present. In this paper, the active learning algorithms are introduced by putting a particular emphasis on the sampling strategies. The iterative processes of the learning engine and the sampling engine are described in detail. The existing theories of active learning are summarized. The recent work and the development of active learning are discussed, including their approaches and corresponding sampling strategies. Firstly, the active learning algorithms are categorized into three main classes according to different ways of selecting the examples. And then, the sampling strategies are summarized by analyzing their correlations. The advantages and the shortcomings of sampling strategies are discussed and compared deeply within real applications. Finally the open problems which are still remained, and the interests of active learning in future research are forecasted.