Candidate Generation for Meal Recommendation System

In this paper we overview a recommendation system architecture proposed by Google and adjust it to for our meal recommendation system. Then we conduct an overview on models used for building recommendation systems, as well as a candidate generation for the recommendation system. We then focus on Wide and Deep model as a main model for our candidate generation stage. This choice is driven by its ability to effectively capture both memorization and generalization aspects, providing a well-rounded solution for the diverse nature of meal preferences. To test and validate our choice we then experiment involving training and testing Wide and Deep model. The model is evaluated both in isolation and in comparison, to widely used Singular Value Decomposition algorithm for recommendation system. Our experimental results show the advantages of Wide and Deep model for candidate generation stage of meal recommendation system.