Batch mode active learning via adaptive criteria weights

• Published in: Applied intelligence (Dordrecht, Netherlands) Vol. 51; no. 6; pp. 3475 - 3489
• Main Authors: Li, Hao, Wang, Yongli, Li, Yanchao, Xiao, Gang, Hu, Peng, Zhao, Ruxin
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.06.2021; Springer Nature B.V
• Subjects: Active learning; Adaptive learning; Artificial Intelligence; Classifiers; Computer Science; Criteria; Learning; Machines; Manufacturing; Mechanical Engineering; Processes; Training; Batch mode active learning; Classification; Adaptive criteria weights
• ISSN: 0924-669X; 1573-7497
• DOI: 10.1007/s10489-020-01953-4

Batch mode active learning (BMAL) is absorbed in training reliable classifier with deficient labeled examples by efficiently querying the most valuable unlabeled examples for supervision. In particular, BMAL always selects examples based on the decent-designed criteria, such as (un)certainty and representativeness , etc. However, existing BMAL approaches make a naive trade-off between the criteria and simply combine them with fixed weights, which may yield suboptimal batch selection since the criteria of unlabeled examples would fluctuate after retraining classifier with the newly augmented training set as the learning of classifier progresses. Instead, the weights of the criteria should be assigned properly. To overcome this problem, this paper proposes a novel A daptive C riteria W eights active learning method, abbreviated ACW, which dynamically combines the example selection criteria together to select critical examples for semi-supervised classification. Concretely, we first assign an initial value to each criterion weight, then the current optimal batch is picked from unlabeled pool. Thereafter, the criteria weights are learned and adjusted adaptively by minimizing the objective function with the selected batch at each round. To the best of our knowledge, this work is the first attempt to explore adaptive criteria weights in the context of active learning. The superiority of ACW against the existing state-of-the-art BMAL approaches has also been validated by extensive experimental results on widely used datasets.