A cost-sensitive semi-supervised learning model based on uncertainty

Aiming at reducing the total cost in cost-sensitive learning, this paper introduces a semi-supervised learning model based on uncertainty of sample outputs. Its central idea is (1) to categorize the samples which are not in training set into several groups based on the uncertainty-magnitude of their outputs, (2) to add the group of samples which have the least uncertainty together with their predicted labels in the original training set, and (3) to retain a new classifier for total cost reduction. The ratio of costs between classes and its impact on learning system improvement is discussed. Theoretical analysis and experimental demonstration show that the model can effectively improve the performance of a cost-sensitive learning algorithm for a certain type of classifiers.