A Hybrid Active Sampling Algorithm for Imbalanced Learning

Class imbalance is a challenging problem in many real-world applications. Various improved algorithms based on oversampling or under-sampling have been proposed, however only gain little when the class distribution is significantly imbalanced. In this paper, we propose a new hybrid sampling algorithm called ALHS, which incorporates two different active learning techniques into oversampling and under-sampling for constructing balanced dataset. Specifically, ALHS under-samples the majority class by actively selecting representative instances in an unsupervised fashion, as a result, the most informative majority instances and a few minority seed points are preserved. For oversampling, we formulate it as a semi-supervised learning problem, and sample possible minority instances from the unlabeled set with label propagation process. By combining the two sampling processes, we can learn from significantly imbalanced training data with less labelling cost. Extensive experiments demonstrate that our algorithm is effective and outperforms other widely used methods in several evaluation metrics.