RICGraph: A Recipe-Ingredient-Compound Graph for Estimating the Nutritional Value of Recipes

Understanding people's health conditions is essential in modern society, and diet management, which is directly linked to understanding health conditions, is one of the most critical issues. Calculating the nutritional value of meals is essential for diet management, but it is not easy. It requires specialist knowledge of the nutrients contained in each ingredient. If we could teach computers to learn about food, there would be a glimmer of hope for achieving this, but no method has yet been established. In this study, we propose a method for estimating the nutritional value of each recipe by constructing a heterogeneous graph using data from multiple data resources. Our heterogeneous graph (RICGraph) is built based on the relationships between the recipe's ingredients, the nutrients contained in those ingredients, and the compounds that make up those nutrients. Therefore, it can consider the relationships between the recipe and the compounds. By performing the learning process according to that structure, it is possible to understand which compounds affect the nutritional value of which recipes and accurately estimate the nutritional value. We estimate the values of the nutrients in the recipes, including protein, fat, saturated fatty acids, carbohydrates, sodium, and energy, and evaluate the accuracy of the estimates. Comparative experiments with multiple LLMs of different types demonstrate the effectiveness of the proposed method.