Yummly allows users to filter recipes by holiday, course, cuisine, and several other attributes. We are working hard to develop additional recipe filters, the latest being one for Super Bowl recipes. What might surprise you is that our recipe understanding system automatically infers recipe attributes. In this post, we’ll talk about some of the signals that are used to predict that a recipe is perfect for the Super Bowl. We’ll learn that Super Bowl food is not ideal for vegetarians and people with New Year’s resolutions, but great for people who love wings, beef, beer, and Mexican food (well, at least Mexican-American food).
Recipe search data
Certain dishes appear more often in searches during particular times of the year. We can use this information to discover associations between dishes and holidays. Here are some dishes that occurred in a much higher percentage of searches on February 5th 2012 (Super Bowl Sunday last year) than 2012 as a whole.
For example, dip was six times more likely to appear in searches on Super Bowl Sunday than usual. Here are some specific dips that were especially popular.
Recipe tags from other sites
Yummly aggregates recipes from across the web, and in some cases these recipes come with tags. However, external tags are treated as evidence, rather than ground truth, in our system. Tag coverage on other sites is often low (e.g., there may be many Super Bowl recipes that are not tagged as such), and the tags that are provided are often incorrect. But we can still learn a lot from them. For example, here are probabilities of ingredients (abbreviated ing) and words in the recipe title for all of the recipes on our site, and for recipes that are externally tagged as Super Bowl.
This analysis tells us that chili and pizza are also common Super Bowl dishes, and that Super Bowl recipes are much more likely to contain beer, beef, and hot sauce.
Other recipe attributes
We infer several attributes for each recipe. In some cases the presence of one attribute increases the likelihood of another. For example, we might expect that Appetizers are more likely to be Super Bowl dishes. Let’s look at the data.
Recipes tagged as Appetizers are more than twice as likely in the set of recipes tagged Super Bowl. In contrast, recipes tagged Salad are three times less likely in the set of recipes tagged Super Bowl.
What about cuisines?
More Super Bowl recipes are Mexican than usual, while fewer are Italian. We saw a few Mexican (Mexican-American) dishes in the search data analysis above.
Whether a recipe is appropriate for a particular diet can also provide information about its attributes. It turns out that Super Bowl food is not very vegetarian friendly: 56.0% of the recipes on our site are vegetarian, but only 37.6% of Super Bowl recipes are.
How unhealthy is Super Bowl food?
Between the wings, beer, and lack of salad, you might get the impression that Super Bowl food is pretty unhealthy. While the data shows that indeed the calories and fat per serving are higher* for Super Bowl recipes, the absolute difference is not especially large. (So you only need to feel a little bit guilty.)
Super Bowl spread- what’s on the table this year?
These are just some of the indicators we can use to automatically identify winning recipes for the Super Bowl. What are some other attributes or indicators that you think would reveal intriguing information about Super Bowl food?
Check back here on Monday when we’ll take a look at this year’s Super Bowl searches. Enjoy the game! And (more importantly) the food!
*The differences are statistically significantly with p < 0.0001 using a t-test.
Photo credits: Chow