BSIL: A Brain Storm-Based Framework for Imbalanced Text Classification

All neural networks are not always effective in processing imbalanced datasets when dealing with text classification due to most of them designed under a balanced assumption. In this paper, we present a novel framework named BSIL to improve the capability of neural networks in imbalanced text classification built on brain storm optimization (BSO). With our framework BSIL, the simulation of human brainstorming process of BSO can sample imbalanced datasets in a reasonable way. Firstly, we present an approach to generate multiple relatively balanced subsets of an imbalanced dataset by applying scrambling segmentation and global random sampling in BSIL. Secondly, we introduce a parallel method to train a classifier for a subset efficiently. Finally, we propose a decision-making layer to accept “suggestions” of all classifiers in order to achieve the most reliable prediction result. The experimental results show that BSIL associated with CNN, RNN and Self-attention model can performs better than those models in imbalanced text classification.