Active learning based on belief functions

Active learning involves selecting a few critical unlabeled samples for manual and credible labeling to improve the performance of the current classifier. The critical step of active learning is the sample selection strategy. Uncertainty sampling is a well-known sample selection strategy, which involves selecting the samples for which the current classifier is uncertain. For the generalized linear model, these samples are usually distributed around the current classification hyperplane. However, uncertain samples include samples near the current classification hyperplane, and samples far from the current classification hyperplane and the labeled samples. Traditional uncertainty sampling fails to describe the latter, and traditional methods are easily affected by outliers. In this paper, belief functions are used to describe the uncertainty that exists in various samples. Furthermore, we propose a sample selection strategy based on belief functions. Experimental results based on benchmark datasets show that the proposed approach outperforms several classical methods. Through this approach, higher classification accuracy can be achieved using the same number of new labeled samples.