摘要
High dimensionality and multicollinearity are increasingly critical for parameter estimation in multivariate analysis to empower intelligent manufacturing, in which severe over-inflated coefficients may degrade the interpretability for modeling and prediction. Partial least squares (PLS) has been employed to handle collinearity for supervised data, in which the latent variables or principal components include all the collinear variables into the derived model. Although the PLS coefficients are decoupled for each collinear variable, the coefficients may be difficult to appropriately reflect their real relationships with the outcome variables. For semiconductor manufacturing, since some variables with collinearity are essential to explain the outcome variables and thus should be kept for further analysis. To fill the gaps, this study aims to develop an integrated approach with domain knowledge for reconstructing the PLS loadings and reinforce the parameter estimation via an innovative approach based on collinear multivariate model (CMVM) by utilizing the ramp function for the loadings of latent variables to optimize the essential coefficients, especially for those coefficients that should be nonnegative in practice. An empirical study was conducted for validation that has shown practical viability of the developed approach to suggest the optimal magnitude of coefficients for each of collinear variables and significantly improve the coefficient interpretation with managerial implications. Indeed, the developed solution is implemented in real settings.