Variance-constrained multi-view ensemble broad network for imbalanced data.

Abstract

In the domain of data classification, the performance of classic broad learning systems (BLS) is negatively affected by significant disparities in data distributions. This leads to a bias towards the majority classes, similar to other conventional techniques. In order to overcome this intrinsic limitation, this study introduces variance-constrained multi-view ensemble broad network (VMEB). First, VMEB utilizes the minimum class variance and class-specific regularization strategies to construct a scatter matrix for each class, thereby enhancing the model's expressive ability for minority classes. Subsequently, we introduce an enhanced kernel mapping technique to address the uncertain random mapping issue in the BLS. This technique generates deterministic hidden nodes through the linear combination of Gaussian functions and arc-cosine functions, enabling more robust feature representation. Furthermore, to overcome the generalization limitations of a single classifier, VMEB embeds this enhanced kernel model into a multi-view ensemble framework. This framework employs a feature rotation strategy to generate diverse views, forcing each base classifier to learn from different perspectives, and achieves highly robust final decisions via majority voting. Extensive experiments conducted on 25 imbalanced datasets demonstrate that VMEB achieves an average AUC improvement of 4.35% compared to baseline. These results confirm the effectiveness of VMEB in handling imbalanced data classification tasks.