Active Class Discovery by Querying Pairwise Label Homogeneity

Abstract

Active learning traditionally focuses on labeling the most informative instances for some well defined learning tasks with known class labels, and a labeler is provided to label each queried instance. In an extreme case, the whole active learning task may start without any available information about the tasks, for instance, no labeled data are available at the initial stage and the labeler is incapable of providing the ground truth to each queried instance. In this paper, we propose an active class discovery method for the case where no randomly labeled instances exist to kick-off the learning circle and the labeler only has weak knowledge to answer whether a pair of instances belong to the same class or not. To roughly identify the classes in the data, a Minimum Spanning Tree based query strategy is employed to discover a number of classes from unlabeled data. Experiments and comparisons demonstrate superior performance of the proposed method for class discovery tasks.