Recommending recipes using time-horizon based user ingredient pool

A non-transitory computer readable storage medium comprising instructions which, when executed by a processor, cause the processor to perform the steps of:

• maintaining, at a recipe recommendation system, a recipe database comprising a plurality of recipes, each recipe including an ingredients list comprising a set of items required to prepare the recipe;
• determining, based on a purchase history of a user and an expected shelf life of one or more items in the purchase history, a user pantry model, the user pantry model comprising a set of items inferred to be available to the user;
• determining, from the plurality of recipes, a set of candidate recipes by, for each of the plurality of recipes:
• identifying, based on a comparison of the ingredients list for the recipe with the user pantry model, a set of missing ingredients comprising one or more items in the ingredients list not included in the user pantry model, and
• selecting the recipe as a candidate recipe based on the set of missing ingredients;
• scoring each of the set of candidate recipes by applying a machine learning model to the candidate recipe, where the machine learning model is trained to generate a score of recommending the candidate recipe to the user, wherein the score indicates a probability of one or more user actions occurring if the candidate recipe is recommended, the one or more actions comprising adding a missing ingredient of the candidate recipe to a current purchase order;
• selecting, based on the scored candidate recipes, one or more candidate recipes to recommend to the user;
• and
• sending, from the recipe recommendation system to a user device associated with the user, the one or more selected candidate recipes for display to the user.