Pool-Based Active Classification Based on Expected Error Reduction with Uncertainty Subsampling

Active learning is a practical machine learning approach useful when labelling data is costly. This study focuses on pool-based active learning, where instances considered most useful for model training are selected from an unlabelled pool set. A well-known label query strategy in pool-based active learning is expected error reduction (EER), which labels instances that are expected to minimise future errors in the model. However, the computational cost of EER grows significantly with the size of the pool set, limiting its practical applicability. To address this issue, we propose to combine EER with uncertainty subsampling, a method that selects queries based on expected error after prefiltering candidate instances using uncertainty. By reducing the size of the unlabelled set for which the expected error is computed, this approach reduces the computational burden of EER while still selecting useful instances. Our experiments show that EER with uncertainty subsampling requires less labelling to achieve a certain performance level than random subsampling. Furthermore, the proposed method allows for a more effective selection of samples that reduce the future error of the model, which is in line with the goal of EER.