#FoodGoals is the first recommendation system that I built as a data scientist. I built it because I had to: all Insight Data Science Fellows are required to complete a data science project. I might have designed another project but this particular topic solved an issue I myself had: I wanted to help consumers make faster decisions when trying to find groceries that are aligned with their nutritional #FoodGoals.
As a consumer myself, I have spent too much time scrutinizing ingredients to make sure they suit my nutritional needs. Moreover, with the rising popularity of online grocery shopping (yes, I realized even in 2019 that this trend was here to stay), such a project could be easily integrated into grocery platforms.
The How
I went searching through the internet for ingredient listings for groceries. I found the USDA’s trove of branded food products and ingredients (and nutrient!) data. This was exactly what I needed to get started. I utilized pandas for data cleaning and characterizing nutrient density of products within their appropriate food categories.
I then became familiar with natural language processing (NLP) techniques as I combed through ingredient listings, removing uninformative words (e.g. salt, flavor, concentrate) and drilling down to base ingredients even when other multi-ingredient items were listed (e.g. if ketchup was listed as an ingredient in a sauce, I would extract tomato, sugar, vinegar and ignore ketchup).
After this, I utilized word2vec—a neural network that learns associations from the co-occurrence of words in a sentence (or in my case, ingredient lists) together, and projects those words into vector space. Thus, this neural network can find related ingredients—and related grocery items based on those ingredients. Finally, I determined the nearness of individual product vectors to each other by calculating cosine similarity.
Having studied nutrition, I knew individuals frequently changed their nutritional goals but I started with a standard slate of options for the potential consumer—for any given product, one could search for recommendations on a low calorie, low carb, low sugar, high protein or low fat version.
The What
And that is how #FoodGoals emerged. This application allows a consumer to request recommendations for grocery items that are similar to one whose ingredient profile they like. While this app is no longer being hosted live, you may view the slides and video of how it operated.