Blending Query Strategy of Active Learning for Imbalanced Data

When the data is imbalanced, often observed in the real-world, important minor class instances that are conducive to accurately predicting the decision boundary are less likely to be queried in the active learning for classification task. Therefore, mitigating the effect of the imbalance in learning is necessary for achieving better generalization. For the alleviation of this problem, this paper considers an active learning algorithm referred to as the blending minority preferential (BMP). The BMP systematically adapts to blend conventional query strategies with the proposed minority preferential queries on a randomly sampled pool dataset. The proposed minority preferential queries are conditioned on an unlabeled data instance predicted to belong to the minor class. A multi-armed bandit is involved in the blending of the two different types of queries for achieving high accuracy and balanced learning in obtaining each query set. The performance of the BMP is validated on datasets that include three imbalanced datasets having binary labels, eleven small structured datasets, modified Fashion-MNIST, CIFAR10 dataets, and two real datasets of the PlantSC and HAM10000. The comparison result shows that the BMP can mitigate imbalanced learning and achieves higher accuracy combined with higher G-mean, compared with conventional queries or other query-based imbalanced active learning.