Cooking With Big Data? Let AI Create Your Next Recipe

Can food and big data cook up a tasty collaboration?

You bet your biscuits they can.

It used to be that this sort of creative work — developing new recipes, discovering new flavor combinations — were the sole purview of human chefs. But computer scientists are proving that AI algorithms can be creative, too.

Computers have been able to combine ingredients in random ways for many years, but that doesn’t mean that a chocolate, ketchup, Brussels sprouts sandwich is a good idea. So how do you move beyond random combinations to creative ones?

In fact, there’s a lot of food science involved. The computer must know, for example, what ratios flour, butter, and leavening most often appear in cookies or cakes in order to create a recipe that will produce the desired result.

Researchers with IBM’s Watson started attacking the problem by using natural language processing (NLP) to process the many, many recipes available on the web, and create a sort of relationship map of ingredients, quantities, and processes associated with different types of recipes. They researched using Wikipedia and other sources to uncover which ingredients are associated with which regional cuisines, and even researched the chemical compounds of foods and which chemical combinations humans rate as being most “pleasant” tasting. It uses what they call the “food pairing principle” to determine what will taste good, based on the idea that ingredients which pair well together will share certain flavor molecules.

With this huge reference database of chemistry, cuisine, and culinary preferences, Watson can generate millions of new recipe ideas that match any preferences input into the algorithm at the start. But which ones will be winners?

That’s where the IBM engineers got particularly clever. Watson then ranks each idea based on novelty and quality. Novelty is defined as how widely does the idea diverge from our ideas of what food “should” be like. (Most people would agree that chocolate and Brussels sprouts is pretty crazy, which would be a very high novelty factor.)

It then also ranks it based on quality, in this case, smell. By analyzing the chemical properties of the different foods in combination, the computer can predict whether or not the dish will smell good — and therefore taste good.

The Watson team have now partnered with food magazine Bon Appetit to improve the process even further, by providing Watson with a cleaner data set. The original experiment was run with creative commons recipes, not all of which had been tested or vetted. In the Bon Appetit database, on the other hand, each recipe has been thoroughly tested in the magazine’s test kitchen, making it a much more reliable source of data.

Other food companies seem likely to jump on this bandwagon and come up with their own applications for big data in the food industry. Restaurant chains like McDonald’s already use data to predict and analyze trends on a store-by-store basis to understand which products are most popular; soon, they may be using data to create new menu items based on data around what has been popular in the past or in a particular region.

Website offers the home chef, restaurant chef, or bartender an opportunity to try the technology for yourself. Anyone can try the Foodpairing inspiration tool for free, which will suggest new combinations based on a main ingredient you define. The tool then suggests best matches based on aroma and flavor profile for foods and drinks. Although the free version only includes 50 foods and 50 drinks, it’s fun to play with to get an idea of how it works.

While big data has already proved itself useful to food companies in every aspect from farm to warehouse to table, these new creative algorithms are out to prove it can even help create better menus.