Improved fine-grained component-conditional class labeling with active learning

We have recently introduced new generative semisupervised mixtures with more fine-grained class label generation mechanisms than previous methods [11], [12]. Our models combine advantages of semisupervised mixtures, which achieve label extrapolation over a component, and nearest-neighbor (NN)/nearest- prototype (NP) classification, which achieves accurate classification in the vicinity of labeled samples. Our models are advantageous when within-component class proportions are not constant over the feature space region "owned by" a component. In this paper, we develop an active learning extension of our fine-grained labeling methods. We propose two new uncertainty sampling methods in comparison with traditional entropy-based uncertainty sampling. Our experiments on a number of UC Irvine data sets show that the proposed active learning methods improve classification accuracy more than standard entropybased active learning. The proposed methods are particularly advantageous when the labeled percentage is small. We also extend our semisupervised method to allow variable weighting on labeled and unlabeled data likelihood terms. This approach is shown to outperform previous weighting schemes.