<p>Recipe plays a vital role in people’s daily life, and the purpose of recipe recommendation is to recommend recipes that users are interested in. Compared with general recommendation tasks, recipe recommendations have the following challenges: (1) Recipe recommendations are non-atomic, and the relation between ingredients is crucial for recipe representation. (2) Different users have different selection criteria for their favorite recipes, and need to capture their personalized preferences. However, existing approaches are limited to the connection between users and recipes, ignoring the importance of ingredients and other contents to recipes, resulting in recommendations that cannot meet users’ personalized preferences and precision requirements. To solve this problem, we propose a new Recipe Multi-Graph Hierarchical Attention Network. The recipe and ingredient graphs are firstly constructed to capture the relationships between recipes and their ingredients. Attentive propagation is used to combine node feature information and structural information to obtain high-quality embedding representation. We integrate attentional mechanisms into the modeling of recipes and ingredients, using attention networks to differentiate the importance of ingredients and adaptively fuse recipe embeddings. The next step is to use hierarchical attention networks to obtain fine-grained user preferences for different recipes and ingredients. Finally, we introduced the recipe knowledge graph for the purpose of enriching the corpus information in order to improve the quality of representation learning. The result is to recommend recipes to users that match their personalized preferences. We conduct experiments on two real datasets and the results show that our approach is state-of-the-art in predicting users’ personalized preferences for recipes.</p>