Aiding food security and sustainability efforts through graph neural network-based consumer food ingredient detection and substitution.

Understanding precisely what is in food products is not always straightforward due to food fraud, differing labelling regulations, naming inconsistencies and the hierarchical nature of ingredients. Despite this, the need to detect and substitute ingredients in consumer food products is far-reaching. The cultivation and production of many ingredients is unsustainable, and can lead to widespread deforestation and biodiversity loss. Understanding the presence and replaceability of these ingredients is an important step in reducing their use. Furthermore, certain ingredients are critical to consumer food products, and identifying these ingredients and evaluating supply-chain resilience in the event of losing access to them is vital for food security analysis. To address these issues, we first present a novel machine learning approach for detecting the presence of unlabelled ingredients. We then characterise the unsolved problem of proposing viable food substitutions as a directed link prediction task and solve it with a graph neural network (GNN).