Abstract
An idea is a collection of existing concepts or words. What makes an idea original or appealing is how these concepts or words are combined in the context in which they appear. Similarly, a food recipe is a combination of ingredients, and it is often evaluated based on how these ingredients fit together to form the whole. In this research, we leverage representation learning methods, specifically word embeddings, to measure the fit among ingredients in the recipe and capture the possibly complex interactions between these ingredients. Using a large-scale online recipe dataset with over 57K recipes, we investigate how the fit between the ingredients relates to recipe popularity (trial) and favorability (ratings). Counter to prior research on creativity, which primarily suggests that creativity is mostly associated with positive outcomes, we find that recipes with unique ingredients have lower trial, but higher ratings given trial. We also find that high fit among ingredients promotes both trial and ratings. We use these findings to develop a generative recipe tool that suggests recipe improvements by adding, removing, or substituting ingredients (http://recipecreativity.com/). We validate our proposed recipe tool using online panel experiments.