Learning recipe ingredient space using generative probabilistic models

In this paper, we demonstrate preliminary experiments using generative probabilistic models on recipe data. Recipes are reduced to lists of ingredients and analyzed in a bag-of-words fashion. We first visualize the highly-dimensional ingredient space and map it to different world cuisines. Latent Dirichlet Allocation (LDA) and Deep Belief Networks (DBN) are then used to learn generative models of ingredient distributions and produce some novel ingredient combinations. First results demonstrate the feasibility of the approach and point to its promise in recipe improvization.