Pool-Based Active Learning with Query Construction

Abstract

Active learning is an important method for solving data scarcity problem in machine learning, and most research work of active learning are pool-based. However, this type of active learning is easily affected by pool size, and makes performance improvement of classifier slow. A novel active learning with constructing queries based pool is proposed. Each iteration the training process first chooses representative instance from pool predefined, then employs climbing algorithm to construct instance to label which best represents the original unlabeled set. It makes each queried instance more representative than any instance in the pool. Compared with the original pool based method and a state-of-the-art active learning with constructing queries directly, the new method makes the prediction error rate of classifier drop more fast, and improves the performance of active learning classifier.