Residual Broad Learning System with Variational Autoencoder for Robust Regression

The Broad Learning System (BLS) has achieved remarkable success in classification and regression problems. Nevertheless, the performance of most BLS models may degrade when dealing with complex nonlinear relationships and contaminated data due to their reliance on single mapping functions and sensitivity to noise through least squares methods. In this paper, we propose a model called Residual Broad Learning System with Variational Autoencoder (RBLS-VAE) to better capture nonlinear relationships and achieve effective denoising. Specifically, residuals are first incorporated into the original features to construct an augmented feature set, where the additional information provided by the residuals complements the patterns not captured in the original features and enriches the representation capability of the input data. And then Variational Autoencoder (VAE) is introduced to better capture complex nonlinear relationships, automatically generate latent representations, and effectively perform data reduction and denoising. Experimental results demonstrate that the proposed RBLS-VAE outperforms traditional BLS and other BLS-based models across multiple datasets, validating its effectiveness and robustness.