The Broad Learning System (BLS) constructs and trains shallow neural networks efficiently by randomly mapping input data to generate feature and enhancement nodes, utilizing ridge regression for weight calculation. Despite its advantages, BLS encounters challenges such as data redundancy and limited adaptability to nonlinear data structures. To address these issues, BLM_DFE was developed, incorporating Kernel Principal Component Analysis (KPCA) for dimensionality reduction and to mitigate redundancy. Nonetheless, there remains potential for improving ridge regression and classification performance. This study introduces a broad learning model for classification with dual feature extraction and discriminative group-sparsity constraints. Our model employs dual KPCA and enhances ridge regression by integrating relaxed labels and group sparsity. This approach enlarges inter-class margins while reducing intra-class margins, preserving the geometric structure of the data for more precise classification. Experimental results demonstrate that our model achieves superior classification accuracy on benchmark databases, surpassing traditional classifiers, the standard BLS, and BLM_DFE. For example, it achieved an accuracy of 85.60% on the GT database, compared to BLM_DFE’s 83.74%. These results highlight its efficacy in handling complex classification tasks.