Difference based query strategies in active learning

In this paper two active learning methods are proposed in the machine learning literature, both of them based on difference calculation idea. One of the new methods is difference sampling query strategy. This strategy calculates a novel difference list and the elements of this list are then able to influence the uncertainty measure of the appropriate unlabelled instance. By taking the ratio of these measures a new informativeness metric is defined, and the aim of the difference sampling strategy is to minimize this ratio. Besides that, expected difference change query strategy was developed using a new metric, the global difference metric for each step. This strategy combines expected model change and uncertainty sampling strategies by taking the expectation of the difference of uncertainty values. The aim of this combined strategy is to query the instance that will most likely result the greatest change in global difference of the next step. The experimental results on image dataset show that both of the difference based sampling query strategies surpass the competitive methods in literature.