Construction of Ingredient Embedding Considering Both Cooking Recipes and Their Ingredients

One of the research methods of food computing that utilizes large amounts of recipe data is the construction of food ingredient embeddings. Ingredient embedding is expected to be applied to a wide range of food-related applications because it enables computers to handle the characteristics of food ingredients. In the existing method of contract ingredient embedding, it was calculated using a graph structure of ingredients and compounds created based on the co-occurrence probability of ingredients in each recipe. However, this method must consider which ingredients are used in which recipes, and whether a correct ingredient vector can be constructed is questionable. In this study, we propose a method for constructing an ingredient embedding by explicitly representing in a graph whether or not each ingredient is included in any recipe. The difference from existing studies is that we construct a heterogeneous graph that includes ingredients, compounds, and recipes. The evaluation experiments showed that the food ingredient classification task using the food ingredient vectors constructed by our method could classify food ingredients with up to 45.8 points higher accuracy than existing methods.