poolbased

Pool-based active learning is a framework in machine learning for efficient labeling, where a model selects which instances from a large pool of unlabeled data should be labeled by an oracle or human annotator. The approach assumes that labels are costly, while unlabeled data is abundant. The typical workflow starts with a small labeled dataset, a large unlabeled pool, and a learning algorithm. At each iteration, a query strategy identifies a subset of unlabeled instances to be labeled, these labeled examples are added to the training set, and the model is retrained. This loop continues until labeling resources are exhausted or a satisfactory performance is reached. Pool-based methods contrast with other settings that either label data on demand or operate on streaming data.

Common query strategies include uncertainty sampling (selecting instances where the model is least confident), margin or entropy-based methods, and more sophisticated approaches like query by committee (disagreement among multiple models). Other strategies incorporate learnable criteria such as expected model change, expected error reduction, or density-weighted methods that prefer informative and representative samples. Batch-mode pool-based active learning selects multiple instances per iteration to improve efficiency in settings where labeling is conducted in rounds.

Advantages of pool-based approaches include making efficient use of labeling budgets and being applicable to large, diverse data pools typical in text, image, and speech tasks. Challenges involve potential sampling bias if the pool is not representative, computational cost of evaluating many candidates, and dependence on the quality of the unlabeled data and the oracle’s labeling.

Pool-based active learning remains a foundational technique for reducing annotation effort while aiming to preserve or improve predictive performance across various domains.