Coupling importance sampling neural network for imbalanced data classification with multi-level learning bias

Imbalanced data classification is a classic challenge in neural network learning. Current rebalancing methods for neural networks mainly rely on resampling and reweighting to alleviate the learning bias caused by imbalanced data learning. It is difficult to properly balance the learning and resampling and rarely dives into the problem of multi-level learning bias. In this paper, we propose a coupling importance sampling to couple the resampling and neural networks learning, while handling both class-level and cluster-level learning biases that occur between and within classes, respectively. Specifically, in the coupling of resampling and learning, as for the class-level learning bias, we extend the resampling from sample to cluster. A composite importance factor is developed to sample important clusters to balance samples between classes. Here, a distribution preservation strategy is additionally maintained to reduce the loss of important samples from resampled clusters. Regarding cluster-level learning bias, a learning regulatory factor is designed to highlight the importance of sampled clusters and avoid the recurrence of imbalance within classes. The proposed method is validated on 34 imbalanced datasets with imbalance ratios ranging from 16.90 to 100.14. The tested results show promising classification performance and prove the advantages of considering the multi-level learning bias in imbalanced data classification.