Speaker: Johnathan Pagnutti, University of California, Santa Cruz
Abstract: Generative methods for recipes have been stymied by a lack of models for the human perception of flavor. Without such a model, generative systems struggle to judge their output based on how humans may actually experience it. In some respects, expert systems have such a model baked in a priori: an expert chef or menu designer has some internal concept of how things taste when they build dishes, and an expert system reflects this knowledge.
Using a dataset of smoothie ingredients with labels for how sweet, bitter, sour, savory and salty a particular recipe is, we aim to learn a model that can predict, for an arbitrary set of ingredients, flavor scores. Using a bag-of-words encoding and reducing the feature space with a Principal Component Analysis, we present some preliminary models using decision trees, random forest and linear regression. This preliminary work showcases a need for a parsing pre-step, and input space that takes not only the presence of an ingredient, but also the amount of it in a particular recipe to try and get better results.