A holistic global-local stochastic configuration network modeling framework with antinoise awareness for efficient semi-supervised regression.

Stochastic configuration network (SCN) has become a sound model due to its universal approximation property with the adaptive configuration of hidden node parameters under a supervision mechanism. However, the basic SCN is essentially a global optimization model and omits the local data structure information, which can be mined by investigating abundant unlabeled samples. In order to address this problem, a holistic SCN modeling framework involving global and local constraints (GLSCN) is presented for effective semi-supervised regression development. For one thing, the model empirical error and the L2 regularization term are assigned from the viewpoint of global optimization. For another, two different local regularization constraints are designed where manifold regularization is utilized to construct local nearest neighbor graphs between samples for taking full advantage of unlabeled samples, and consistency regularization constraint is introduced to further handle the potential random noise. By injecting random noise into the unlabeled samples, local consistency regularization can enforce the prediction consistency effect of unlabeled samples and improve the local smoothness of each sample further. In this way, not only can a large amount of unlabeled information be fully exploited, but the model robustness is also effectively enhanced. Lastly, the validity of the developed GLSCN is demonstrated by two industrial cases of a sulfur recovery unit and a continuous stirred tank reactor. [ABSTRACT FROM AUTHOR]