Saito,P.T.M., de Rezende,P.J., Falcao,A.X., Suzuki,C.T.N., Gomes,J.F.
Statistical analysis and pattern recognition have become a daunting endeavor in face of the enormous amount of information in datasets that have continually been made available. In view of the unfeasibility of complete manual annotation, one seeks active learning methods for data organization, selection and prioritization that could help the user to label the samples. These methods, however, classify and reorganize the entire dataset at each iteration, and as the datasets grow, they become blatantly inefficient from the user's point of view. In this work, we propose an active learning paradigm which considerably reduces the non-annotated dataset into a small set of relevant samples for learning. During active learning, random samples are selected from this small learning set and the user annotates only the misclassified ones. A training set with new labeled samples increases at each iteration and improves the classifier for the next one. When the user is satisfied, the classifier can be used to annotate the rest of the dataset. To illustrate the effectiveness of this paradigm, we developed an instance based on the optimum path forest (OPF) classifier, while relying on clustering and classification for the learning process. By using this method, we were able to iteratively generate classifiers that improve quickly, to require few iterations, and to attain high accuracy while keeping user involvement to a minimum. We also show that the method provides better accuracies on unseen test sets with less user involvement than a baseline approach based on the OPF classifier and random selection of training samples from the entire dataset