Feature fusion-based weighted broad learning system for imbalanced data

Broad Learning System (BLS), characterized by its flat architecture as an efficient neural network, has garnered significant interest for its benefits in terms of training velocity and network scalability. However, BLS does not perform well when dealing with imbalanced classification problems. First, conventional BLS neglects the intrinsic logical information contained in the original input data, which may be critical for classification and prediction. Second, BLS fails to emphasize the significance of minority classes in imbalanced data scenarios. In this paper, we propose a feature fusion-based weighted broad learning system. Unlike traditional BLS, we combine the original features and broad features as fused features for training. Additionally, we introduce a weight generation mechanism that assigns higher weights to minority classes, enhancing the model’s focus on these classes. A closed solution is available for the issue, and the transformation matrix can be effectively resolved. The experimental results on multiple imbalanced datasets show that our method is superior to other imbalanced methods.