Active Zero-Shot Learning

In multi-label classification in the big data age, the number of classes can be in thousands, and obtaining sufficient training data for each class is infeasible. Zero-shot learning aims at predicting a large number of unseen classes using only labeled data from a small set of classes and external knowledge about class relations. However, previous zero-shot learning models passively accept labeled data collected beforehand, relinquishing the opportunity to select the proper set of classes to inquire labeled data and optimize the performance of unseen class prediction. To resolve this issue, we propose an active class selection strategy to intelligently query labeled data for a parsimonious set of informative classes. We demonstrate two desirable probabilistic properties of the proposed method that can facilitate unseen classes prediction. Experiments on 4 text datasets demonstrate that the active zero-shot learning algorithm is superior to a wide spectrum of baselines. We indicate promising future directions at the end of this paper.