FIM-Based Pairwise Selection for Active Learning on Imbalanced Datasets

Active learning has exhibited strong ability in improving efficiency of practical classification tasks. However, when applied to imbalanced datasets, traditional selecting strategies for active learning can be severely disturbed by redundant majority instances thus fail to offer inherent capability. In this paper, Fisher Information Matrix based pair wise selection (FIMPS) is proposed to solve the malfunction of active learning on imbalanced datasets. The FIMPS is primarily derived from FIM-based selection by introducing pair wise selecting strategy to adapt to the features of imbalanced data distribution. During active learning, the FIMPS is performed on an initial classifier generated by SVM variant for imbalanced problem to extract a balanced dataset from unlabeled pool for training set update. Thus, the retraining process can be freed from the interference of imbalanced problem. FIMPS is compared with various selecting approaches on a variety of real-world datasets. The result shows that the proposed method outperforms other strategies in enhancing classification performance of active learning.