Determining sensory drivers of complex metadescriptors through regression modelling

Abstract

In sensory science, terms such as creaminess often lack precise definitions due to their multi-modal nature. Least absolute shrinkage and selection operator (LASSO), a regression technique known for automatic predictor selection, and partial least squares regression, which handles multicollinearity, were compared for their ability to accurately identify the underlying sensory attributes driving creaminess perception. Twenty-eight sensory attributes were selected after discussions with milk consumers. Thirty-two milk samples were chosen to represent these attributes, spanning a wide range of creaminess. Quantitative descriptive analysis, with trained panellists, and a consumer study (n = 117 New Zealand milk drinkers) assessed the sensory attributes and creaminess ratings, respectively. LASSO and PLSR were compared for their predictive ability and attributes retained using sensory attributes (trained panel) as predictors and creaminess ratings (consumers) as the response variable. LASSO identified four key sensory attributes with a good model fit (R2 = 0.951), while PLSR suggested thirteen (R2 = 0.933). LASSO is effective in uncovering pertinent attributes within a complex sensory experience enabling cost-effective research. PLSR offers a comprehensive model for extensive product development. This research provides an alternative approach for determining pertinent attributes in complex metadesciptors. Resulting models offer clearer targets for product development, thus increased commercial gains.