Task-aware graph convolutional network for active learning

Active learning to select the most valuable examples for annotation from unlabeled data pools in high-dimensional datasets is particularly challenging and has attracted considerable attention. However, the impact of annotated examples on global data was often not considered. In this paper, we propose a novel batch active learning strategy called TA-GCNAL based on the perspective of task model. To consider data distribution in the labeled and unlabeled pool, we first calculate the correlation between data through a graph convolution network to predict the example's uncertainty. Then, a subset of examples with high uncertainty is generated to replace the unlabeled data pools. It transformed into a probability space based on the task model for core dataset selection and annotation. Finally, we compute the representativeness of the core dataset in probability space for the example batch selection, taking into account the correlation between labeled and unlabeled pools. Our proposed TA-GCNAL achieve superior performance than the state-of-the-art methods on various benchmark datasets. The improvement is particularly significant on unbalanced and large number of classes datasets.