Partial least squares with a regularized weight

Abstract

The key issue in partial least squares (PLS) is the computation of loading weight which describes the relations between the responses and observation variables. In this paper, we propose a simple modified PLS regression method to adaptively compute the loading weight. The weight is solved according to a regularization framework with a similarity constraint to keep the spectra–concentrate relations and meanwhile to penalty the variables with a large distance to the response. The regularized weights can accurately reflect the spectra–concentrate relations, and lead to a good-fitting and parsimonious model which is less sensitive to the latent variables. The proposed method is compared, in terms of root mean square errors of prediction, correlation coefficient and cumulative explained variance, to PLS on simulated and real near-infrared spectra data sets. Experimental results demonstrate the efficacy and effectiveness of the proposed method.