Interpretable Relational Representations for Food Ingredient Recommendation Systems
Abstract
Supporting chefs with ingredient recommender systems to create new recipes is challenging, as good ingredient combinations depend on many factors like taste, smell, cuisine style, texture, chef’s preference and many more. Useful machine learning models do need to be accurate but importantly– especially for food professionals – interpretable and customizable for ideation. To address these issues, we propose the Interpretable Relational Representation Model (IRRM). The main component of the model is a key-value memory network to represent the relationships of ingredients. The IRRM can learn relational representations over a memory network that integrates an external knowledge base- this allow chefs to inspect why certain ingredient pairings are suggested. Our training procedure can integrate ideas from chefs as scoring rules into the IRRM. We analyze the trained model by comparing rule-base pairing algorithms. The results demonstrate IRRM’s potential for supporting creative new recipe ideation.