A variable selection method based on multicollinearity reduction for food origin traceability identification

In spectral modelling analysis, multicollinearity problems among the spectral variables are prevalent, which may reduce the accuracy of the analysis result. To reduce the effect of multicollinearity between variables in classification analysis, a new strategy of variable selection named as multicollinearity reduction-based variable selection (MR-based VS) is proposed. Characteristic variables were selected based on inter-class significant difference and intra-class correlation evaluation, which reduced data multicollinearity and ensured the selected variables were more relevant to the categories. It was combined with supervised pattern recognition methods of least squares discrimination analysis (PLS-DA) and uncorrelated linear discriminant analysis (ULDA) for the identification of the red wine and olive oil from different geographical origins. The results show that compared with the full-spectrum model and the traditional successive projection algorithm (SPA) variable screening model, the MR-based VS strategy reduces the multicollinearity between variables while ensuring the maximum difference among the different classes, as a result, it obtained the superior classification results. Therefore, MR-based VS can effectively extract categorical features, eliminate redundant information, and improve model interpretability, which shows potential for enhancing the ability of the spectral qualitative analysis model in different fields.