Search

Abstract: A method for product optimization based on external preference mapping (EPM) is proposed and tested on a publically available data set. As with EPM an individual’s preferences are modeled using latent variables (LVs) derived from trained panelists’ sensory evaluations. The method differs from traditional EPM. First, it is advised that more LVs be used in modeling with the hope of improving model fit. Second it is suggested that a robust criterion be used in model selection and culling in order to achieve “better” models. Finally, standard, sophisticated optimization approaches are advised for estimating product optima. The method gives an estimate of the optimal product profile in the original sensory space for a given consumer group. A number of comparisons are made, e.g. modeling with LVs derived from partial least squares (PLS) regression vs. those derived from PCA, scaling vs. not scaling preference data, model selection with a genetic algorithm vs. backward selection, modeling with 1st order terms vs. 1st and 2nd order terms and optimization using two different sets of optimization constraints. Given enough latent variables, the method retained a large proportion of the individuals in the given population and fit preference data well in both a calibration and a validation sense. However, evaluating a previously published optimum with the new method underestimated a validated result. When 6, 7, 8 and 9 LVs were used, the associated optima were relatively consistent within a given approach. [Copyright &y& Elsevier]

Abstract: A relatively new regression method, L-partial least squares regression (L-PLSR), is used to model aggregate liking for 17 varieties of fresh tomatoes. Four missing value estimation methods were evaluated and the “winner” applied to the hedonic data. Cluster analysis was performed on the actual and estimated hedonics to arrive at three preference clusters. Usage and Attitude (UA) profiles for the clusters plus descriptive data for the tomato varieties were used in combination in L-PLSR to model the cluster preference profiles. The results were compared to partial least squares regression (PLSR) models, a principal components analysis (PCA) model and chi-square goodness-of-fit analysis of the UA data. Model selection using a genetic algorithm and backward selection is noted. [Copyright &y& Elsevier]