Without giving too many details, Joe Roberts (@1winedude) revealed the other day that he has been helping the folks at VineSleuth with their new app for wine drinkers. He mentioned that the app (called wine4.me) will be available “soon” and that it will use data from wine tasters to do stuff (with science). But what kind of stuff?
According to VineSleuth’s website, the app will do quite a few things, but its main focus will be wine recommendations based on past user feedback. A kind of virtual wine sommelier.
Does this sound like something familiar? It should; the Pandora radio service tries to do the same thing with music. It starts with a familiar song or artist that the user likes and then it suggests similar songs based on data about the songs that are available. Wine4me appears to be after the same goal. Knowing this, a few people have been wondering:
How will wine4.me work? (probably)
It starts with data. In order to find similarities between wines, each wine must be characterized. This is where Mr. Joe Roberts comes in. In order to characterize the wines, someone has to taste them—shocking, but true.
This tasting, however, is not the usual setup. Wines are not rated by overall quality. Instead, each wine is rated on many different characteristics. As a reference point, Pandora claims to use ~400 characteristics per song in order to deliver music recommendations. In their system, each song is listened to for about 30 minutes, and, one by one, the listener ranks the magnitude of each characteristic. A similar setup will most likely be used for gathering the wine characteristic data for the wine4.me app. In a song these characteristics are things such as tempo, key, and the gender of the singer. For wine, just imagine every adjective you’ve ever heard attributed to a wine and then visualize someone giving a number for each of these (and more) when ranking a wine.
The reason it is done this way is to eliminate subjectivity. The preferences of the rankers don’t affect the results when the characteristics are ranked individually. All that matters is that the rankers define a certain wine as having the correct amount of each characteristic. While I don’t really agree with their claim to novelty, this is how VineSleuth’s website describes it:
“VineSleuth has invented a revolutionary method for wine evaluation that ensures our characterizations are objective, repeatable, and predictable. And yes, that’s rare.”
As an oversimplified example, let’s say that a user indicates to app that they like 2 wines that happen to both be dessert wines. Giving these wines as a starting point (assuming that these wines have been “characterized”), the app would then compare each characteristic to find what matches. It would see that wine “A” had a sweetness value of 9.5 and that wine “B” had a sweetness value of 9.0. From there, it would calculate that this small difference shows that sweetness is the common characteristic. The app then would suggest to the user another wine with a similar sweetness value.
In this example, we only looked at one basic characteristic, but what if you like a wine who’s characteristics are balanced? This too can be calculated, assuming that the important characteristics are ranked. Even subtle relationships can be discovered through machine learning techniques (that are far too boring to go into). Suffice it to say that if the data is good then the recommendations will be good.
On top of the “characteristic data” driven recommendations, an app like this could also use user ratings (“likes” and “dislikes”) to help each other. You’ve probably seen this already, as well. The phrase “Other people who bought this also bought _____” is a common sight in e-commerce these days, and there is no reason why wine4.me wouldn’t leverage this same method of recommendation.
What’s the catch?
So far this sounds like a good deal. After all, if Pandora can suggest music fairly well then why not wine? There are a few unique problems for wine characterization, however.
One difference is that music is much easier (and cheaper!) to characterize than wine. Listening to the same song a hundred times (for more accurate data) only costs the time of the person listening. Good luck trying that with wine! Each taste costs money, even if the wine is free (someone’s gotta clean the glasses), and there are only so many wines that can be tasted before getting fatigued.
Another issue is that wine changes over time. While most would agree that aging is a good thing, it obviously causes some problems for anyone wanting to characterize wine. Without constant updates, aged wine would not be accurately characterized, leading to broken expectations.
What about wine faults? Wine faults could label normally good wine as a “dislike”. This would steer the learning algorithm in the wrong direction, causing future disappointment.
There are some good differences, too
One good difference is that the qualifications for “good data” might be relaxed for wine. After all, taste is the weakest of the 5 main senses. This means that instead of spending 30 minutes to characterize a single song, as Pandora’s raters (reportedly) do, much less time than this could probably be spent on each wine.
Another thing that wine4.me won’t have to worry about is royalty fees. When Pandora listeners log into the streaming service, Pandora has to pay fees for the right to broadcast. These fees simply wouldn’t apply for wine4.me’s service.
Also, as more users use wine4.me, more data is collected about what people like. This data is gold. Using this “like” and “dislike” data, VineSleuth can show wine makers what the average drinker drinks. For a price, of course. User data like this would also be able to show in what regions certain characteristics are appreciated most, allowing the wine trade a new way to gauge interest. Larger wine producers could use this data to target certain segments of the wine market, customizing their offerings based on the preference data that VineSleuth provided.
This brings me to the biggest difference between wine and music. Compared to wine, music is tiny. Music sales last year were at $7.1 billion (U.S. sales according to RIAA), which admittedly is down from its peak of double that in 1999. When compared to wine, this is nothing; $34.6 billion was sold in wine inside the U.S. last year. Wine may not be consumed by everyone, but it is worth much more.
Ultimately, there are way too many variables to say whether such a scheme will work out for sure. Obviously, VineSleuth has run the numbers, and they feel confident in their ability to reach profit, or they wouldn’t do it.
As I understand it, they would need a certain number of active users to provide useful, valuable data to the wine industry. Things may be hard until they reach this “critical mass” if they reach it at all. Once they do reach this, though, they will only get bigger and better than before. Then, when our grandkids grow up, we’ll hear them say, “Hey, check out this old radio service I found. They suggest music just like wine4.me suggests wine!”