Phase I: Consumer assessment
In this phase, the methodology proposed by Camargo et al.,30 was adapted to identify and qualify the importance and the interactions among consumer preferences. To this purpose, a set of semantic attributes, which accurately describes the perception of a group of users, allows calculating the fuzzy measures. This method employs the backpropagation algorithm with constraints proposed by Grabisch34,35 to approximate the fuzzy measure accurately, using the Choquet integral.36 This algorithm is efficient when training data are limited, requires little computing time, and a low memory cost.29 As the input range is not strictly defined in this study, the learning is combined here with the output through a linear error regression, which ensures accurate decision and better convergence.37 The details of the definitions and computations of Choquet integral and fuzzy measures are presented in Section A of the Supplementary material.