Robust Broad Learning System for Uncertain Data Modeling

Junwei Jin, C. L. Philip Chen, Yanting Li

Abstract: Broad Learning System (BLS) has achieved good performance in classification and regression problems, and the computational efficiency is especially outstanding. However, there exists various outliers or noise in the sampling data, which puts a robust requirement on the algorithms. Standard BLS is sensitive to the contaminated data because of its composition structure. In this paper, we propose a robust version of BLS called RBLS to improve its generalization on contaminated data modeling. In RBLS, the ℓ2-norm based cost function will be replaced by ℓ1-norm style cost function. The Augmented Lagrange Multiplier (ALM) method is applied to optimize the new model iteratively. The experiments on function approximation and real-world regression demonstrated that the RBLS method has better modeling performance for sampling data with outliers or noise.