useR! 2024

The sensory profiles of coffee products can be generated by using predictive models with inputs from specific electrodes immersed in the liquid coffee. The electrical signals are generated by molecules diffusing through selective bio-polymers and reaching electrodes. These signals are time-series data, and various features are calculated from them (such as transient peaks or steady-state averages). These features are then used to train models for the prediction of sensory profiles or proximity to a given reference product. Six features are calculated for each of the 15 pairs of electrodes resulting in 90 variables for the classification model (proximity to a reference) or the regression model (sensory profile). These variables can be correlated to a certain extent and lead to over-fitting or unnecessary recording. This work presents a method for reducing the number of variables to a minimum relevant set using clustering and random forest variable importance. It shows that a proper selection results in a more robust model across different experimental batches.