How Tastry “Taught a Computer How to Taste.”

How Tastry uses novel chemistry and AI to predict consumer preferences.

From the outset, the question we wanted to answer was: “Can we decode the unique flavor matrices of sensory-based products, and the unique biological preferences of consumers to accurately predict likability?”  The short answer is yes.

However, early in our research we found that existing chemical analysis methods and existing consumer preference data, provided statistically insignificant correlations or predictions. We knew we would have to create our own data in order to make progress.

First, we needed to create an analytical chemistry method that would provide as much transparency to the chemistry as possible (including volatiles, non-volatiles, dissolved, spectral data, and so on.)  We also needed to decode the flavor matrix in a way that could be translated to help approximate how humans experience that chemistry on their palate.

Second, we needed to create a method to constantly and accurately obtain, augment, and track the biological sensory preferences of a large, diverse, and ever growing group of actual consumers to serve as our ground truth.

Why current methods fail to predict consumer preference for sensory-based products

When we started our research in 2015, we had the hypothesis that everything you need to know about the flavor of wine, that is to say the taste, aroma, texture, and color – exists in the chemistry. However, what was missing was a more comprehensive method of analysis.

To explain this limitation, it is important to understand that the chemistry of sensory-based products is largely focused on quality control, i.e., how much of this analyte is in that mixture? The focus is not typically to evaluate all the analytes, their relative ratios, or how they combine on the human palate to create flavor. This is the blind spot we needed to illuminate because there are dynamic interactions taking place among hundreds of compounds on a human palate. A human palate experiences a “chemical soup” of flavor compounds at the same time, not one compound at a time like a machine does. The interactions between these multiple compounds in combination with the unique biology of each consumer, provide critical context as to what features of the chemistry are expressed to that person.

To the extent that sensory is taken into account, simply put, the typical approach looks like this:

• Survey data shows that people like butter.
• Diacetyl is a compound typically associated with the flavor of butter.
• If we make a chardonnay with more diacetyl, more people will like it.

Core problems with this approach.

1. Flavor cannot be predicted by quantification of compounds alone. A given concentration of diacetyl may be perceived as butter in one wine or vintage, but not in another. This is because there are hundreds of other compounds in the wine, and depending on their concentrations and ratios, diacetyl could either be masked or expressed. Unlike a machine, humans are experiencing all the compounds at once, their senses are not analyzing each compound individually, therefore any individual given quantification is not necessarily predictive.

 

2. Humans perceive and communicate flavors differently. Even among a panel of experts, half the experts may describe something as tasting like apple, and the other half may describe it as pear. And the average consumer is even less predictable. From our research, we don’t believe that human taste is sufficiently tangible to be accurately communicated simply through language from one person to another. Our descriptors are too vague, and our definitions vary based on individual biology and cultural experiences. For example, in the U.S. most consumers describe the perception of benzaldehyde as “cherry”, but most consumers in Europe describe it as “marzipan”…even in the same wine.

 

3. The flavors consumers perceive have no correlation with whether or not they actually like it. In our research it is observed that consumers do not decide to purchase a wine because it tastes like cherry. They simply make the judgment that they liked the wine, and they are likely to like it again.

Example: This lack of understanding is not unique to the wine segment. We have met with executives and researchers at some of the largest flavor and fragrance companies in the world. One executive described his frustration with a recent project to create a new lavender chocolate. This company spent millions of dollars seating and running focus groups with consumers who specifically loved chocolate, loved lavender, and loved lavender chocolate. Ultimately the results were that the respondents agreed it was lavender chocolate, but that they also agreed they didn’t like that particular lavender chocolate.

As a result of these insights, we concluded that we should focus our research on predicting what chemistry matrices consumers liked, and to what extent, as opposed to what flavors they perceive.

How Our Approach is Different

Garbage-in, Garbage-out. When it comes to data quality, we realized a valid training set could not be generated from existing commercial or crowd-sourced data. We would have to create our own, in-house.

The first thing we needed was a chemistry method that would provide visibility on the delicate balance of the volatile, nonvolatile, dissolved solids, spectral data, etc., of a wine in one snapshot, to be more relatable to the human palate.

Years of experimentation resulted in a methodology that generates over 1 million data points per sample. This granular and overwhelming amount of data is then processed by machine learning algorithms that were designed by our data science team to decode the interdependencies which inform human perception based upon the ratios of the analytes and groups of analytes.

Once we had proven efficacy for this method, we began analyzing and decoding the flavor matrix of many thousands of wines worldwide and have since developed a comprehensive flavor matrix database of the world of wine.

Relating Consumer Preferences to Chemistry

Next, we had to understand what flavor matrices various consumers preferred by having them taste and rate the wine we had analyzed. Over the years we have run regular double-blind tasting panels with thousands of consumers, each tasting many dozens or hundreds of wines over time.  Respondents include newcomers to wine, typical wine drinkers, experts, winemakers, and sommeliers.

Crowd-sourced systems typically miss or ignore critical data. For example, on the Parker scale, most people won’t even score a wine below the mid-80pt. range.  But we’ve learned that consumers dislike what they dislike, more than like what they like.  Therefore, it’s critical to have a full picture of preference – especially negative preferences.

We used our novel machine learning to understand the consumers unique preferences for various types of flavor matrices in the wine. Over time, this allowed us to accurately predict their preferences for wines they had yet to taste.  During this process, we also learned that individual wines, as well as individual preferences, are almost fingerprint-like in their uniqueness.  We concluded that, contrary to customary industry practices, consumers and wines cannot be accurately grouped, or collaboratively filtered, into generalizations.

Example:  Two females can share the same geography, culture, ethnicity, education, income, car, phone, and both love Kim Crawford Sauvignon Blanc; but one can love Morning Fog chardonnay and the other can hate it.  The only reliable predictive visibility rests with their biological palate.

How to scale this innovation? 

What we had created was great, but tasting panels are expensive and time consuming. It would be impossible to run an annual tasting panel of all 248 million Americans over the age of 21 to understand what wines they will like.

We wanted to design a scalable tool that had the same efficacy in predicting a consumer’s preferences, without requiring participation in tasting panels or expressing their preferences for a large set of previously tasted wines.

Our solution was to have the AI select simple food items which shared aspects of their chemistry with wines in an assortment.  Respondents in our tasting panels answered several hundred such questions about their preferences for foods and flavors that are not directly related to wine; such as, “How do you feel about green bell pepper?”, or “How do you feel about mushrooms?”

These questions were used by TastryAI as analogs to the types, and ratios, of compounds commonly found in the underlying chemistry of wine.  As humans, we cannot decipher or understand these complex correlations and patterns, but as it happens teasing out these complicated relationships is an excellent problem for machine learning to solve.

With this data, TastryAI learned how to predict a consumer’s preference for wine, based on their answers to the Food Preference Survey. What resulted was our ability to eliminate the need for any wine specific data from a consumer to predict their preference for wine.

How much data do we need to understand consumer preference?

Although we started with hundreds of food preference questions, the more that are answered the more accurate the results, there are diminishing returns after 9-12.  With the Pareto principle at work, the best performing food preference questions provided approximately. 80% understanding of a consumer’s palate.

As of today, there is typically a 10-12 question survey for red wine, and another 10-12 question survey for white, rosé, and sparkling wine.

This allowed a scalable solution. Since we launched in various pilots years ago, there are now many similar whimsical-looking quizzes on ecommerce sites.  A consumer takes a 30-second quiz about whether or not they like blackberries or coffee, and they are rewarded with wine recommendations. The difference is that those quizzes are at most tasting note filters, i.e., if you like blackberries you’ll like a wine described by someone as tasting like dark fruit, or if you like coffee then you’ll like a wine described by someone as being astringent.  But we have learned that if those descriptions are accurate for that person’s palate, it has no predictive power as to whether or not they will like the wine; but it is engaging, consumers like quizzes.

Tastry’s recommendations are tied to the flavor matrix of the wine. TastryAI is not a tasting note filter, it isn’t asking if you like the aroma or taste of mushrooms in your wine, it’s trying to understand the ratios of compounds you like or don’t like based on your biological palate preferences.  Each question provides many layers of insight because each question overlaps and feeds into other questions.  So, after asking about mushrooms, perhaps the next question is “How do you feel about the taste of green bell pepper?”  The AI may know that there are, for example, 33 compounds in a given ratio generally responsible for the perception of mushrooms, and 22 compounds generally responsible for the taste of green bell pepper – but importantly some of those compounds exist in both.  If you say you love mushrooms, but hate green bell pepper, then the AI is more confident you like some compounds, more confident you dislike other compounds, and those that overlap are likely contextual.

So, you can kind-of imagine a multidimensional Venn diagram, where the AI is teasing out which compounds you like or dislike in combination with other compounds.

And with this flavor preference survey, and consumer feedback, we collect anonymized palate data from around the World. An e-commerce site, or big box retailer, can launch the Tastry Quiz on the app, and have thousands of responses within hours from consumers across the U.S. The only other data we acquire is a zip code. We use the zip code to apply a derivation of a Bayesian ridge, which takes the geographic distribution of the known consumer palates we collect and monitor, and other data, and predicts the rest of the 200M+ viable consumer palates in the U.S.  We use this enhanced dataset as the source of truth, and to provide predictions on how wines will perform in a market on a store, local, or regional level.

Tastry Virtual Focus Group

Upon analyzing a wine, decoding its flavor matrix, and evaluating its palatability against the combination of actual and virtual palates, the AI is currently 92.8% accurate in predicting the aggregate U.S. consumer rating for the wine. In other words, the AI can predict the average 5-star rating for a wine within +/- 1/10^th of a star.

It is easiest to think of the AI as a “Virtual Focus Group” of consumer preferences.

Wineries use TastryAI to run simulations on how consumers will perceive their wine, even before they invest years and millions of dollars into making it.  Wholesalers use TastryAI to determine the regions in which various wines will perform best. Retailers use TastryAI to optimize their assortment on the shelves and online. And consumers use TastryAI to avoid the risk of buying a wine that they are not going to like.