Online Passive Aggressive Active Learning and Its Applications

We investigate online active learning techniques for classification tasks in data stream mining applications. Unlike traditional learning approaches (either batch or online learning) that often require to request class label of each incoming instance, online active learning queries only a subset of informative incoming instances to update the classification model, which aims to maximize the classification performance using the minimal human labeling effort during the entire online stream data mining tasks. In this paper, we present a new family of algorithms for online active learning called Passive-Aggressive Active (PAA) learning algorithms by adapting the popular Passive-Aggressive algorithms in an online active learning setting. Unlike the conventional Perceptron-based approach that employs only the misclassified instances for updating the model, the proposed PAA learning algorithms not only use the misclassified instances to update the classifier, but also exploit those correctly classified examples yet with low prediction confidence. We theoretically analyze the mistakes bounds of the proposed algorithms and conduct extensive experiments to examine their empirical performance, in which the encouraging results show clear advantages of our algorithms over the baselines.