Enhancing GISMo: Integrating Recipe Contexts into a Graph-Based Ingredient Substitution Module

To comply with dietary restrictions, individuals seek to replace ingredients in culinary recipes. The advanced model Graph-based Ingredient Substitution Module, called GISMo, recently was proposed to facilitate the learning of ingredient substitutions within recipes. We identify some potential improvements in GISMo’s functionality for generating substitution recommendations based on recipe contexts by capturing various graph representations of recipe contexts, including ingredient-recipe relationships and cooking actions. Furthermore, we introduce a novel benchmark dataset comprising substitutions for specific recipes, validated by four domain experts. To assess the proposed improvements, we conduct experiments on both existing and new substitution datasets. Our findings demonstrate that the proposed changes in GISMo can augment the diversity of food substitution recommendations without compromising prediction quality.