Construction of a model for predicting sensory attributes of cosmetic creams using instrumental parameters based on machine learning

Abstract This study aims to enhance the sensory evaluation of skin creams by using machine learning to predict sensory attributes based on instrumental parameters, addressing the limitations of conventional methods. Extensive instrumental parameters, including rheological, tribological, and textural properties of ten different skin creams, were collected, and 22 sensory attribute scores were obtained from trained expert evaluations. Pearson’s correlation analyses and multiple supervised learning algorithms were applied to establish relationships between each sensory attribute and the instrumental parameters, respectively. The results showed that K-Nearest Neighbors, AdaBoost, and LightGBM were the algorithms with the best performance for most sensory attributes, and the overall model achieved over 95% prediction accuracy for 80% of the sensory dimensions, demonstrating strong reproducibility and accuracy in the verification test, with predicted sensory scores closely aligning with actual values. However, issues such as model overfitting, discrepancies between training and testing data distributions, and increased noise in the testing data, which result in the residuals from the testing set exceeding those from the training set, need to be further addressed in future research. In summary, this data-driven approach offers a more efficient, accurate, and standardized method for sensory evaluations, with the potential to support the development and optimization of skincare products.