Multi-Kernel Broad Learning System Based on Elastic-Net with Random Fourier Features

The Broad Learning System (BLS) features a simple yet efficient network structure, with its core being the fast and random generation of hidden layers; however, this generation method not only fails to effectively capture the nonlinear characteristics in the task, but also generates certain ‘redundant nodes', which can negatively affect its learning capabilities. In this study, we propose an improved version of BLS, named the KEFBLS, aimed at enhancing the feature extraction capability of the hidden layer through the integration of multi-kernel technology and network sparsification strategies, complemented by deeper feature extraction using random Fourier features. the KEFBLS first combines polynomial and wavelet kernels to boost the nonlinear mapping capabilities of data; then, it applies the elastic-net method to refine the BLS objective function, removing low-impact hidden layer nodes to reduce redundancy and create a more streamlined network; finally, KEFBLS employs random Fourier features to map the processed hidden layers, further enhancing the network's feature extraction capabilities, constructing a new learning model. Our experimental results on three UCI regression datasets demonstrate that KEFBLS surpasses other methods in terms of learning efficiency and model performance.