Fuzzy-Rough-Set-Based Active Learning

Determining the informativeness of unlabeled samples is a key issue in active learning. One solution to this is using the sample's inconsistency between conditional features and decision labels. In this paper, a fuzzy-rough-set-based active learning model is proposed to tackle this problem. First, the consistence degree of a labeled sample is computed by the lower approximations in fuzzy rough set, which reflects its minimum membership in the decision class. Then, the concept of sample covering is proposed to measure the relationship between labeled samples and unlabeled samples. Afterward, the memberships of an unlabeled sample belonging to different decision classes are computed based on the covering degrees of labeled samples on it. Finally, these memberships are used to form a sample selection criterion to measure the sample's inconsistency. By applying Gaussian kernel-based similarity relation to the aforementioned processes, a support vector machine (SVM)-based active learning scheme is developed. Experimental results demonstrate the effectiveness of the proposed model.