People who would like to improve their eating patterns can make small changes in their diet by substituting ingredients in recipes. \Good" substitutions may have multiple dimensions and de nitions, and one de nition is to maintain the recipe while adhering to constraints that may include personal preferences, allergies, and nutritional or other dietary considerations. Our proposed system automatically nds and ranks plausible substitution options using a knowledge graph of food information. We evaluate our substitute ranking heuristic using a novel data set of ground-truth substitutions, showing promising preliminary results.
Eating habits can play an important role in improving personal health. For example, patients diagnosed with diabetes typically receive recommendations to monitor their intake of speci c nutrients from foods (like carbohydrates and protein) to treat their condition. If patients nd that some nutrients that they are monitoring need adjustments, they may choose to modify their diet by substituting ingredients in some of their recipes. Such substitutions can remove restricted types of ingredients (e.g., common allergens) or replace ingredients that most negatively a ect patient health (e.g., replacing potatoes to reduce carbohydrate intake). Substituting individual ingredients rather than strictly following a new meal plan allows patients to eat familiar meals while maintaining their dietary goals.
Our work aims to empower people to improve their eating habits by
simplifying the process of identifying ingredient substitutions that satisfy their dietary
constraints. We facilitate this using our FoodKG [
Intuitively, our approach to ingredient substitution ranking should consider similarities in the properties of ingredients as well as the similarity of the recipes in which they are used. We also must consider how to determine whether ingredients that are \similar" are also good substitutes. Some combinations of ingredients (e.g., garlic and olive oil) may be similar in the sense that they co-occur frequently in recipes, but they may not be substitutable. On the other hand, some ingredients that often are good substitutes (e.g., potatoes and cauli ower) may be dissimilar in terms of some properties like their food classi cation.
In order to create a ranked set of substitution options, we seek to develop a
heuristic combining several scores based on latent and explicit semantic
information about ingredients. We use two sources of latent semantics in the form of
word embeddings based on ingredient names. The rst is a Word2Vec [
For explicit semantic information, we compute two substitutability scores
based on the intuitions that (1) good substitutions should pair well with
similar ingredients, and (2) good substitutions should be used in similar recipes.
Additionally, we use the linked ontology of food from FoodOn [
Ground-Truth Data Collection Because there is no widely accepted goldstandard data set for ingredient substitutions, we have taken the approach of selecting an online source to serve as a \ground-truth" for a wide variety of ingredient substitutions. We collected substitution data from The Cook's Thesaurus3, yielding a set of 2,300 substitutions pairs (i.e., pairs of \target ingredient" to \substitute ingredient"). 1,161 unique ingredients were present in the data set, providing substitutions for 928 target ingredients. 3
We evaluate our results by producing a list of ranked substitution options for each ingredient and comparing it against our ground-truth data. Since our method produces substitution rankings for each target ingredient, we frame our approach as an information retrieval problem and compute mean average precision (MAP) and mean reciprocal rank (MRR) as our evaluation metrics. MAP allows us to assess our algorithm's ability to rank all substitutions because the number of correct substitutes varied for each ingredient (e.g., in our evaluation data set, \potato" had 6 substitutes while \white asparagus" only had 1). MRR then gives us insight into the highest-ranked correct substitution for each ingredient. With our ltering strategy applied, our combined heuristic achieves a MAP of 0.418 and MRR of 0.499 over our ground-truth data. Comparing it against a simple baseline using word embedding similarity, our heuristic shows some promising improvement over the Word2Vec model (MAP 0.297 and MRR 0.385) and SpaCy (MAP 0.274 and MRR 0.371). 3 foodsubs.com
Although our approach was able to show better performance than simple baselines, there are many limitations in the methods and evaluation. Our ltering strategy was not always able to successfully lter out undesirable options because of incorrect links to FoodOn classes as well as limitations of the information captured in FoodOn. Sometimes the valid substitutions from the algorithm that were ltered out were not in line with our ground-truth data. For example, garlic powder is classi ed as a subclass of garlic in FoodOn, which led our approach to incorrectly lter it out of garlic's substitution options. This re ects our reuse of an ontology that perhaps had di erent intentions for how its subclass relationships may be used from our reuse expectations.
Although relying on a single website as our ground-truth allowed our evaluation to avoid subjectivity issues that may arise from user studies, it is also limited in its scope. While the source of our evaluation data provides a wide variety of substitution options, it does not cover all possible substitutions that people may use (e.g., many websites cite zucchini as a valid substitute for potatoes, but it was not in our ground-truth data). Finally, our current evaluation method does not assess other factors that may be involved in judging the goodness of substitutions (e.g., the tness of a substitution for a particular type of diet, or the tness of a substitute in vastly di erent types of recipes). 5
Using FoodKG and its linked information about ingredient classi cation, our approach for identifying ingredient substitutions shows promise when compared against simple baselines. Our current best heuristic achieves a MAP of 0.418 and MRR of 0.499 in our collected ground-truth substitution data set. Work remains to improve both our strategies for ingredient substitute ranking as well as its evaluation. Moving forward, we are also expanding our approach to include more explicit semantics about the \healthiness" of substitutions in combination with more sources to develop a more comprehensive evaluation.
This work is partially supported by IBM Research AI through the AI Horizons Network.