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

From the outset, the query we desired 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.

Nonetheless, early in our research we found that existing chemical evaluation methods and existing consumer preference data, provided statistically insignificant correlations or predictions. We knew we might need to create our own data with a view to make progress.

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

Second, we wanted to create a technique to continuously and accurately obtain, augment, and track the biological sensory preferences of a big, diverse, and ever growing group of actual consumers to function our ground truth.

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

After we began our research in 2015, we had the hypothesis that every part you’ll want to know in regards to the flavor of wine, that’s to say the taste, aroma, texture, and color – exists within the chemistry. Nonetheless, what was missing was a more comprehensive method of study.

To clarify this limitation, it will be significant to grasp that the chemistry of sensory-based products is basically focused on quality control, i.e., how much of this analyte is in that mixture? The main focus isn’t typically to guage all of the analytes, their relative ratios, or how they mix on the human palate to create flavor. That is the blind spot we wanted to light up because there are dynamic interactions happening amongst a whole bunch of compounds on a human palate. A human palate experiences a “chemical soup” of flavor compounds at the identical time, not one compound at a time like a machine does. The interactions between these multiple compounds together with the unique biology of every consumer, provide critical context as to what features of the chemistry are expressed to that person.

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

• Survey data shows that folks like butter.
• Diacetyl is a compound typically related to the flavour of butter.
• If we make a chardonnay with more diacetyl, more people will prefer it.

Core problems with this approach.

1. Flavor can’t be predicted by quantification of compounds alone. A given concentration of diacetyl could also be perceived as butter in a single wine or vintage, but not in one other. It’s because there are a whole bunch of other compounds within the wine, and depending on their concentrations and ratios, diacetyl could either be masked or expressed. Unlike a machine, humans are experiencing all of the compounds directly, their senses are usually not analyzing each compound individually, due to this fact any individual given quantification isn’t necessarily predictive.

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

3. The flavors consumers perceive haven’t any correlation with whether or not they really prefer it. In our research it’s observed that buyers don’t determine to buy a wine since it tastes like cherry. They simply make the judgment that they liked the wine, they usually are prone to prefer it again.

Example: This lack of information isn’t unique to the wine segment. We now have met with executives and researchers at a few of the largest flavor and fragrance firms on this planet. One executive described his frustration with a recent project to create a brand new lavender chocolate. This company spent thousands and thousands of dollars seating and running focus groups with consumers who specifically loved chocolate, loved lavender, and loved lavender chocolate. Ultimately the outcomes were that the respondents agreed it was lavender chocolate, but that additionally they agreed they didn’t like that individual lavender chocolate.

In consequence of those insights, we concluded that we must always focus our research on predicting what chemistry matrices consumers liked, and to what extent, versus what flavors they perceive.

How Our Approach is Different

Garbage-in, Garbage-out. In terms of data quality, we realized a sound training set couldn’t be generated from existing industrial or crowd-sourced data. We’d need to create our own, in-house.

The very first thing we wanted was a chemistry method that might provide visibility on the fragile balance of the volatile, nonvolatile, dissolved solids, spectral data, etc., of a wine in a single snapshot, to be more relatable to the human palate.

Years of experimentation resulted in a strategy that generates over 1 million data points per sample. This granular and overwhelming amount of information 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 flavour matrix of many hundreds 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 grasp what flavor matrices various consumers preferred by having them taste and rate the wine we had analyzed. Through the years now we have run regular double-blind tasting panels with hundreds of consumers, each tasting many dozens or a whole bunch 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 instance, on the Parker scale, most individuals won’t even rating a wine below the mid-80pt. range.  But we’ve learned that buyers dislike what they dislike, greater than like what they like.  Subsequently, it’s critical to have a full picture of preference – especially negative preferences.

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

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

Tips on how to scale this innovation? 

What we had created was great, but tasting panels are expensive and time consuming. It could be not possible to run an annual tasting panel of all 248 million Americans over the age of 21 to grasp what wines they are going to like.

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

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

These questions were utilized by TastryAI as analogs to the kinds, and ratios, of compounds commonly present in the underlying chemistry of wine.  As humans, we cannot decipher or understand these complex correlations and patterns, but because it happens teasing out these complicated relationships is a wonderful problem for machine learning to resolve.

With this data, TastryAI learned easy methods 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 necessity for any wine specific data from a consumer to predict their preference for wine.

How much data do we want to grasp consumer preference?

Although we began with a whole bunch of food preference questions, the more which are answered the more accurate the outcomes, there are diminishing returns after 9-12.  With the Pareto principle at work, the most effective performing food preference questions provided roughly. 80% understanding of a consumer’s palate.

As of today, there is often a 10-12 query survey for red wine, and one other 10-12 query survey for white, rosé, and sparkling wine.

This allowed a scalable solution. Since we launched in various pilots years ago, there at the moment are many similar whimsical-looking quizzes on ecommerce sites.  A consumer takes a 30-second quiz about whether or not they like blackberries or coffee, they usually are rewarded with wine recommendations. The difference is that those quizzes are at most tasting note filters, i.e., for those who like blackberries you’ll like a wine described by someone as tasting like dark fruit, or for those who like coffee you then’ll like a wine described by someone as being astringent.  But now we have learned that if those descriptions are accurate for that person’s palate, it has r; but it surely is engaging, consumers like quizzes.

Tastry’s recommendations are tied to the flavour matrix of the wine. TastryAI isn’t a tasting note filter, it isn’t asking for those who just like the aroma or taste of mushrooms in your wine, it’s attempting to .  Each query provides many layers of insight because each query overlaps and feeds into other questions.  So, after asking about mushrooms, perhaps the subsequent query is “How do you’re feeling in regards to the taste of green bell pepper?”  The AI may know that there are, for instance, 33 compounds in a given ratio generally answerable for the perception of mushrooms, and 22 compounds generally answerable for the taste of green bell pepper – but importantly a few of those compounds exist in each.  If you happen to say you like mushrooms, but hate green bell pepper, then the AI is more confident you like some compounds, more confident you dislike other compounds, and those who overlap are likely contextual.

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

And with this flavor preference survey, and consumer feedback, we collect anonymized palate data from across the World. An e-commerce site, or big box retailer, can launch the Tastry Quiz on the app, and have hundreds of responses inside hours from consumers across the U.S. The one other data we acquire is a zipper code. We use the zip code to use 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 remaining of the 200M+ viable consumer palates within the U.S.  We use this enhanced dataset because the source of truth, and to offer 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 mixture of actual and virtual palates, the AI is currently 92.8% accurate in predicting the mixture U.S. consumer rating for the wine. In other words, the AI can predict the common 5-star rating for a wine inside +/- 1/10^th of a star.

It’s easiest to think about 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 thousands and thousands of dollars into making it.  Wholesalers use TastryAI to find out the regions by 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 chance of shopping for a wine that they are usually not going to love.