Robust stochastic configuration networks with maximum correntropy criterion for uncertain data regression.

This paper develops a robust stochastic configuration network (RSCN) framework to cope with data modelling problems when the given samples contain noises or outliers. Technically, RSCNs are built by generalizing the objective function used in our original stochastic configuration networks with maximum correntropy criterion (MCC) induced losses (the proposed algorithm is termed as RSC-MCC). The half-quadratic (HQ) technique is employed to optimize the penalty weights for each training sample, aiming to weaken the impacts caused by the noisy data or outliers throughout the training session. Alternating optimization (AO) methodology is used to renew the RSCN model in company with updated penalty weights determined by HQ methods. The performance of RSC-MCC algorithm is compared with some existing methods, such as the probabilistic robust learning algorithm for neural networks with random weights (PRNNRW), RVFL networks, improved RVFL networks (Imp-RVFL), and our recent work RSCNs with kernel density estimation (RSC-KDE), on two synthetic function approximation examples, four benchmark datasets and one educational data modelling case study (for student learning performance prediction). The experimental results show that RSC-MCC performs more favourably in robust data analytics, and further indicate that our proposed RSCN framework (both RSC-KDE and RSC-MCC) has a good potential for real-world applications.