EvoRecipes: A Generative Approach for Evolving Context-Aware Recipes

Generative AI e.g. Large Language Models (LLMs) can be used to generate new recipes. However, LLMs struggle with more complex aspects like recipe semantics and process comprehension. Furthermore, LLMs have limited ability to account for user preferences since they are based on statistical patterns. As a result, these recipes may be invalid. Evolutionary algorithms inspired by the process of natural selection are optimization algorithms that use stochastic operators to generate new solutions. These algorithms can generate large number of solutions from the set of possible solution space. Moreover, these algorithms have the capability to incorporate user preferences in fitness function to generate novel recipes that are more aligned with the fitness objective. In this paper, we propose the
$EvoRecipes$
framework to generate novel recipes. The
$EvoRecipes$
framework utilizes both Genetic Algorithm and generative AI in addition to
$RecipeOn$
ontology, and
$RecipeKG$
knowledge graph. Genetic Algorithm explore the large solution space of encoded recipe solutions and are capable of incorporating user preferences, while LLMs are used to generate recipe text from encoded recipe solutions.
$EvoRecipes$
uses a population of context-aware recipe solutions from the
$RecipeKG$
knowledge graph.
$RecipeKG$
encodes recipes in RDF format using classes and properties as defined in the
$RecipeOn$
ontology. Moreover, to evaluate the alignment of
$EvoRecipe$
generated recipes with multiple intended objectives, we propose a fitness function that incorporates novelty, simplicity, visual appeal, and feasibility. Additionally, to evaluate the quality of the
$EvoRecipe$
generated recipes while considering the subjective nature of recipes, we conducted a survey using multi-dimensional metrics (i.e. contextual, procedural, and novelty). Results show that
$EvoRecipes$
generated recipes are novel, valid and incorporate user preferences.