An imbalanced binary classification method via space mapping using normalizing flows with class discrepancy constraints

For imbalanced binary classification, most methods classify data in the original space by balancing data or modifying classification algorithms. However, for datasets with severe overlaps and complex class boundaries, the classification performance is limited by the ability of the classifier to fit the decision boundary. A practical alternative is to map the data to a latent space with lower classification difficulty. In this paper, we propose an imbalanced binary classification method via space mapping using normalizing flows with class discrepancy constraints. The flow-based model is employed to map the original data to a latent space, ensuring the accuracy of the mapping and the authenticity of the mapped distribution. To solve the problem of overlaps and within-class imbalance, class discrep-ancy constraints, including global and local constraints, are proposed to modify the flow -based model. The former maps different subclusters to the same cluster of the latent space, and the latter increases the separability of different classes. Classification thus can be car-ried out in the latent space with simpler distribution and high separability to achieve bet-ter classification performance. Experimental results on 35 KEEL and UCI public datasets indicate that the proposed method outperforms other state-of-the-art methods on F1 -measure and G-mean. (c) 2022 Elsevier Inc. All rights reserved.