Imbalanced Broad Learning System with label relaxation and sample weight adaptation

The Broad Learning System (BLS), as a lightweight network architecture, has been extensively applied to various classification and regression tasks. However, BLS and its variants remain suboptimal for addressing imbalanced classification problems. These models often pay little attention to the quality of original features. Their supervision mechanisms typically rely on strict binary label matrices, which impose limitations on approximation and fail to align with the underlying data distribution. Additionally, they generally do not differentiate the contributions of majority and minority classes, leading to a bias towards majority classes in predictions. In this paper, we propose a novel imbalanced BLS framework that integrates label relaxation and sample weight adaptation to address challenges in imbalanced classification tasks. First, genetic programming is employed to optimize the original features, improving data representation capability. Then, a latent label space is constructed based on pairwise label relationships, which serves to achieve flexible label relaxation. Furthermore, a dynamic weighting mechanism is proposed based on intra-class and inter-class distributions to balance the influence of majority and minority classes. Extensive experiments conducted on 30 benchmark datasets demonstrate that the proposed method significantly outperforms various state-of-the-art approaches, with average G-mean and AUC scores of 89.6% and 90.0%, respectively. These results validate the effectiveness and superiority of the proposed model in addressing imbalanced classification tasks.