The Broad Learning System (BLS) employs two successive random feature mappings and incorporates ridge regression to construct an efficient classifier. This system achieves performance comparable to Deep Neural Networks, while reducing computational resources. However, BLS exhibits limitations such as a high risk of overfitting, and inadequate robustness to noisy data, which can be addressed through ensemble learning. Nevertheless, traditional ensemble learning models are limited by the lack of interpretability. Therefore, we propose a novel ensemble learning method for BLS with feature selection via Approximate Shapley Value, named Shapley-Value-Based Ensemble Broad Learning System (SE-BLS). Specifically, by analyzing the validation losses of BLS-based weak classifiers, the Approximate Shapley Value is calculated to determine feature contributions and collaborations, which can guide effective feature selection. Based on the reduced features, a regularized classifier is trained for final voting to achieve superior performance on datasets with limited samples and high noise levels. Additionally, a new data weight updating method is introduced for SE-BLS to improve its stability in dealing with imbalanced data. Notably, for image datasets, our method can provide visual analysis of pixel contribution, pixel collaboration, and Class Activation Mapping to enhance interpretability. To validate the performance of the proposed model, we conduct tests on multiple structured and image datasets to assess its classification performance and visualization capabilities. These results are then compared with those of several advanced models. Furthermore, ablation experiments are performed on various aspects of the model's structure to demonstrate its effectiveness.