Περίληψη:
This article documents an approach to predicting positive sensory attributes – fruitiness, bitterness, pungency-of virgin olive oil from its chemical characteristics, using machine learning methods. The dataset used in this study included forty-nine olive oil samples of the Koroneiki variety from nine selected olive mills, evenly distributed in the region of Messinia, Greece. The samples were analyzed for free acidity, peroxide value, the UV absorption for the determination of the extinction coefficients, phenolic compounds (total secoiridoid phenols, oleocanthal, oleacein, oleuropein aglycon and ligstroside aglycon) and sterol compounds (total sterols, cholesterol, campesterol, stigmasterol, d7-stigmasterol, erythrodiol, uvaol, b-sitosterol). Sensory analysis of the samples took place 20–30 days after their sampling date and the intensity of three positive attributes (fruitiness, bitterness and pungency) was measured. The authors used the least absolute shrinkage and selection operator (Lasso) for feature selection and then applied ordinary least squares (OLS) methods to build the final models. Three Python-based forecasting machine learning models for each sensory characteristic (fruitiness, bitterness, and pungency) were built and evaluated in comparison to one another in terms of the performance metrics of root mean squared error (RMSE) and mean absolute percentage error (MAPE), using repeated 5-fold cross-validation. The interacting effects among the sensory features were also considered for developing the two regression models, while the third model was only based on chemical attributes. The results obtained, revealed a significant relationship between each sensory attribute and the intensity of the other two, with the respective prediction models demonstrating a highly satisfactory level of performance. Furthermore, models that employed only chemical indices as predictors provided strong evidence that chemical indices alone were sufficient to predict the intensities of the sensory attributes. The findings of this study establish the predictive value of the constructed models, which might be utilized to support panels in training and calibration. © 2023 The Authors
Συγγραφείς:
Kottaridi, K.
Anna, M.
Vasilis, D.
Aimilia, R.
Vasileios, N.
