The Center for Data Innovation spoke with Michael Nestrud, senior manager of global sensory science and consumer affairs at Ocean Spray. Nestrud discussed how he uses data science to explore consumer preferences and improve meal rations for soldiers.
Joshua New: What is sensory science, and why is it important for companies like Ocean Spray?
Michael Nestrud: Sensory science is an offshoot of psychology and is the science of evoking, measuring, analyzing, and interpreting the human sensory response to stimulus. Now, this is kind of broad, as everything we know about anything comes through our senses! Within the context of Ocean Spray, we’re talking about taste tests and flavor. We deliver food and beverages to consumers—either in a controlled lab environment or ship it to their homes for authentic context—and measure their response using specific types of questionnaires. We mostly operate in the quantitative realm on our team, which means collecting enough data to provide input for data-driven decisions. A given sensory questionnaire could have 35 questions on anywhere from two to four products administered to 350 people, or over 30,000 data points to sift through. We use a combination of descriptive statistics and modeling to find insights, such as which products people like, why they like them, and how to make them better.
New: Part of your job involves analyzing Twitter posts about cranberries. What can you learn from this kind of information?
Nestrud: The Twitter work is an internal hobby at the moment. Unstructured data as everybody knows is quite difficult to work with and I wanted to do more than “sentiment” which I’ve never quite trusted. By first removing noise—all tweets with “Cranberry” and “Zombie,” for example—and then converting words to their base lemmas, you can start to count frequencies, classify words, and look for patterns that way. Graph theory is well suited to exploring this type of data. You can learn, for example, what fruits are most commonly combined with cranberry in recipes—citrus, for example—or cocktails. I like this type of emergent data analysis because it is generally less biased by the experimenter than survey data.
New: Some of your recent research focuses on quantifying “emotional profiles” for use in product research. How does this work? I can’t imagine emotions being easy to quantify and analyze.
Nestrud: There’s a lot of “rich” discussion about the quantification of emotional reaction to foods and beverages. This is usually done with a CATA—Check All That Apply—type question on a sensory survey. The words might be happy, sad, delighted, angry, surprised, guilty, and you can imagine more. There’s several sets of emotion word lists scientifically developed and validated for different purposes. The biggest controversy with measuring emotions via questionnaires is whether or not you’re measuring emotions that people are feeling in that moment, or emotional associations—does a Snickers really make you “satisfied”? Or do you just think it does because that’s what the commercials say? However, what we do know, is that there’s common patterns in the way people respond to these questions, and it does have a certain predictive component for consumer choice behavior.
New: Can you describe how you used graph theory to make better pairings of food items in U.S. Army field rations?
Nestrud: Graph theory made army field rations better by actually looking at not the hedonic value of individual food items, but looked at what pairs of food items were the most important for predicting whether or not an MRE (Meal Ready to Eat) was a cohesive meal. The theory goes that the more it fits the concept of a meal, the more a warfighter will consume it, and nutrition is extremely important in the field. We learned, for example, that if the seasoning packet matched the entrée, and the entrée matched the side item, that the rest didn’t matter so much with respect to whether or not it was considered a cohesive meal. Given that there’s 13 individual components in an MRE—snack, side, hot beverage, cold beverage, bread, spread, dessert, and others—this was a big insight and helped focus development.
New: What has been the most surprising thing you’ve learned from taking a data-driven approach to understanding what influences people’s taste?
Nestrud: The most surprising insight about using data is that it quite often proves personal opinions wrong.The potential success or failure of every decision we make with respect to product characteristics can be evaluated with the data we collect. This is a powerful tool for risk-based decision making. We don’t always get it right—after all if predictions were always correct, developing products would be easy. However, predictions with data have more certainty than opinions and memory, which are fraught with the limitations of human bias and memory.