Distributed representation of tags for Active Zero Shot learning

Extreme multi-labeled classification (XMLC) refers to the problem of tagging items to its most relevant subset of class labels from an extremely large set of labels. Since it is practically difficult to obtain training instances for each of such a large set of labels, the problem is better tackled by a zero-shot approach, where training instances of only a small subset of labels are used. In this paper, we focus on active zero-shot classification, where one can select which subset of labels to obtain training data for. The challenge is to intelligently select a small subset of labels (seen classes) for which to obtain training data, such that the trained model can accurately predict a large number of unseen labels. We present a model named Distributed Representation of Tags for Active Zero-Shot Learning (DiRTAZL), which learns label similarities from an external knowledge source, to intelligently select the seen classes. We experimented on three data sets to demonstrate that our proposed model is superior to prior models for active zero-shot learning paradigm.