Abstract
The work is devoted to solving such a complex scientific and applied problem of dietetics as the formation of recipes for dishes of different food categories. The purpose of the article is to develop a method for recommending recipes based on machine learning, which allows taking into account the personal preferences and dietary restrictions of the user. The proposed method consists of three stages: candidate generation, scoring, and recipe ranking. At the candidate generation stage, a two-tower model using neural networks is used. This model allows for the effective integration of user and dish information into a common vector space and the quick retrieval of relevant recipes using a nearest neighbor search algorithm. After that, during the candidate generation stage, filtering takes place, which weeds out recipes that contain allergens or products restricted by the user. The scoring stage involves the user evaluating the recipes obtained from the candidate generation stage and forming a list of recommendations, which are ranked in order of recommendation in the next stage. For this purpose, machine learning models such as matrix factorization, decision tree-based models, and deep learning are used in the evaluation stage. These machine learning models take into account user preferences. The scoring stage also proposes modules that take into account contextual factors such as viewing history, day of the week, and time available for cooking. The ranking stage is implemented using a maximum marginal relevance algorithm, which ensures a balance between the relevance of recipe suggestions to user preferences and the diversity of recommendations. The scientific novelty of the work lies in the fact that, for the first time, a method has been developed that allows taking into account the user's personal preferences and their preferences regarding ingredients and cooking techniques. The proposed method has high potential for building intelligent personalized meal planning systems that promote a healthy lifestyle for users.

This publication has 8 references indexed in Scilit: