You love the blog, so subscribe to the Beervana Podcast on iTunes or Soundcloud today!

Thursday, February 17, 2011

Beer Genome Project (My Get-Rich-Slow Scheme)

In a post yesterday, Stan reprises a meme I discussed last August, wondering if the methods used to create Pandora's Music Genome Project would work for beer:
Anyway, while I was reading Gladwell’s article — which delves into the subjectivity involved in setting “objective” standards — Pandora managed to feed me song after song that I didn’t feel the need to skip. It’s been a while since The New York Times explained how “The Music Genome Project” works, but it’s still a fascinating story. ... Would it be possible to do something similar for beer?
In a word: yes. But it wouldn't be easy. To create a "music genome," Pandora employed a team of musicologists to break music down into elements and code them. If you've ever done qualitative research, the process would be familiar. They looked for objective, technical features, like major or minor key tonality, rhythm patterns, vocal harmonies and so on. But then they bravely tackled the subjective as well, looking at, to quote that Times piece Stan linked: "To what extent, on a scale of 1 to 5, does melody dominate the composition of 'Hey Jude'? How 'joyful' are the lyrics? How much does the music reflect a gospel influence? And how 'busy' is Stan Getz’s solo in his recording of 'These Foolish Things'? How emotional? How 'motion-inducing'?"

What results are a series of markers that form the song's "genome." I booted up my account and looked at the first song that played. Pandora described why it selected this song:
"We're playing this track because it features rock influences, off-beat style, highly syncopated drum beats, use of modal harmonies, a slow-moving bass line, dominant use of piano riffs, affected synths, an acousti-synthetic sonority, trippy soundscapes, prevalent use of groove, and many other similarities identified in the Music Genome Project."
Note what it doesn't include: genre. This is part of the genius of Pandora. For decades, radio stations have played certain genres of music, conditioning us to partition our tastes off using this criterion. What you find when you tune into Pandora are certain qualities of music that you weren't aware you liked. Apparently I'm big on syncopated drums and complex rhythms. I like electronic influences, but I couldn't care less about vocals. The music Pandora selects for me has an amazingly referential quality. Without realizing it, I was setting up "channels" that were effectively the same channel--the music was related closely enough that I kept getting the same music no matter which band I stared out with. (That's on my indie-based music, anyway.)

A Beer Genome
So how would a beer genome work? Pandora wisely started with musicologists; amateurs like me trying to code music would have missed whole layers of subtle markers. To construct a beer genome, you'd need people with sophisticated palates and a substantial background in beer tasting. From there, I think you'd follow the music method pretty closely. There are a lot of objective qualities in beer--the beer flavor wheel offers a nice template. You'd probably expand it to capture a few more qualities like color and effervescence and perhaps strength.

Then, to capture the essence of a beer's nature beyond mere descriptors, you'd need to come up with a list of subjective qualities, like "sunshiny lightness," or "intense greenness." (And this is where you'd come in for a lot of hell: so many beer geeks and homebrewer types hate anything that isn't quantifiable; they wouldn't just disagree with the categories, they'd disagree with the subjectivity principle.) My guess is that it is these qualities that would help you overcome the conceptual rigidity of style--which, obviously, would need to be abandoned, just like Pandora abandoned music genre.

Maybe you'd come up with a description like this:
"We've selected this beer because it features a light but round straw-colored body, gentle cookie malt sweetness, perfumy, floral hop aroma, fruity notes of citrus and melon, piquant, lively mouthfeel, a clean, crisp finish, sunny brightness, and many other similarities identified in the Beer Genome Project."
(I was trying to describe Sierra Nevada Pale Ale from memory there.) Part of the reason Pandora succeeds is because it has scores (hundreds?) of markers. What seems banal in the particular begins to sort out into pretty small categories. If you like SN Pale, perhaps the qualities would lead you to try a wit, say, which wouldn't be expected if you hewed strictly to style. More to the point, it would guide you not just to other similar styles, but particular beers. And this is where the genius of such a project would lie. If you tune in a radio station, you select one based on the blunt dictates of genre. You might end up enjoying 75% of the music and really liking 25%. But with Pandora, that number goes up quite a bit--I enjoy probably 90% of the music and really like half of it. If you could reliably find a way to navigate the tens of thousands of beers produced every year in the world, the same thing would happen.

In terms of implementing it, well. As Stan says, it's a pipe dream. You'd need to get a team of trained tasters who could assemble a list of say 2000 beers at launch. It would be a great app--you could easily charge a few bucks for it. But the list would have to grow quickly; since beer is regional, you'd need ten thousand or more to ensure that everyone had a decent selection of options. That would require some kind of serious investment. I'd be happy to take some calls from VCs, but I don't expect them.

Ah, but imagine, you're in the mood for something like, oh, say Hair of the Dog Adam. You boot up your Beerdora, hoping to be able to scratch the itch. Adam--that'll be a tough beer to match, you think, but no, there they are, a dozen suggestions just waiting for you.

13 comments:

  1. I like this concept, though I think something at least similar already exists. I'm not sure the methodology is as precise or quantitative as the Music Genome Project, but I recently discovered pintley.com, which seems to (at least try to) accomplish the same goal of recommendations. I haven't populated enough of my personal ratings yet to see how the feedback turns out, but at least a few of the recommendations so far seem to be pretty good.

    ReplyDelete
  2. Pintley uses a different methodology. Essentially, their system is the same that exists through much of the internet. I don't know what it's called, but it uses an associative methodology to extrapolate based on other beers you like. The Pandora methodology actually judges a song based on its own qualities, rather than associations it has with other songs.

    Which is not to say that Pintley won't work. Netflix uses that method and does pretty good at guessing my tastes. It suggested Veronica Mars, of all things, which I loved.

    ReplyDelete
  3. Coupling two of your last three post
    - Beerdora
    - Beer Sensory Science
    together with the
    - Scientific American article referenced in 'Beer Sensory Science'
    suggest Beerdora's task is much more difficult than Pandora's task. Each of your olfactory worlds differ. Ie, may not be on the same planet much less the same continent.

    I particularly enjoyed these post, thanks.

    ReplyDelete
  4. Bret - As far as Pintley having the same goal, yes and no. Yes, in that we do personalized recommendations. However, no, we don't do it like this post talks about.

    Jeff - You're right, Pintley uses collaborative filtering. And I like to think that we do work ;)

    Depending on what the markers are, a beer Genome project might only end up recommending truly similar beers.

    Let's think about how Pandora actually works. You start by picking an artist or a song to create a station. Pandora then starts playing songs from that artist, and then branches out to similar songs from other artists. So, let's say we pick a specific IPA for a beer "station". I would be pretty disappointed if a beer genome started recommending other beers that taste very similar to that IPA, since the markers are the same or similar. That kind of defeated the purpose of a beer recommendation engine, which is to recommend *different* things to try, not more of the same, right?

    So a straight "port" of Pandora definitely wouldn't work. However, if a beer genome database is built, perhaps something interesting could be done with that data.

    ReplyDelete
  5. But why do you need highly trained beer 'geeks' (tastologists?) to do this? Why can't you use the masses, or at least the more beer educated masses? Sites like Beer Advocate have a lot of people who know their beer quite well. If you were able to piggyback on a site like that which has a higher level of beer connoisseur, you can have lots of people judge certain qualities of the beers. You'd be surprised how accurate you could get by using a higher quantity of slightly less 'experienced' talent.

    ReplyDelete
  6. I don't know I find that when I listen to Pandora... which is a significant amount of time because it encompass my 8 to 12 hour brew days... that at some point the music takes a turn towards a different "genre". Usually I find that some odd song gets played and then all the following songs take note and eventually I get Johnny Cash when I was originally listing to Little Wayne. In fact you could listen to any station and a little Johnny will kick in eventually. I'm pretty sure its some inside joke at Pandora. (BTW I like Johnny Cash, but he gets way to much time on Pandora, does he own stock?)

    On another note couldn't you use brewery press releases/tasting notes to categorize the beers. That way you wouldn't have to pay "trained" tasters (I wish I was a trained taster). Because in reality its all perception when tasting a beer so why not let the breweries do the leg work?

    Cheers, I love the idea of Beerdora!
    Kyle Larsen
    Double Mountain Brewery

    ReplyDelete
  7. Pintley also seems to quietly pay for its self by suggesting that I might want to try every single beer that Sam Adams makes, when pretty much all of the beers I have rated highly are sours or CDA's. I know the people working there have stated that they have a secret revenue stream, but between people getting more points for "loving" a certain silent advertisers beers, and Pintley's software suggesting the beers from other silent advertisers, it is pretty obvious how they are paying their bills.

    I would be more apt to find a site that isn't quite this transparent, that charges a small annual fee, which is what i do with Pandora.

    Don't get me wrong, I love their site, but for me it is mostly so I can easily remember what I drank, and if I liked it or not.

    Whatever method works the best to make recommendations, it needs to be paid for at some point, especially if the "inventor" wants to one day be a millionaire :P

    Now, start a summit to outline all of the genomes, and I will gladly meet at a bar and drink while scribbling ideas on the back of a napkin.

    ReplyDelete
  8. I really found the topic of subjectivity to be interesting. I totally agree that the average beer geek would balk at subjective/abstract descriptions of a beer, but I also think beer drinkers experience beer on both the objective and subjective levels and that isn't something that should be suppressed.

    So much of my experience with a beer is tied to the context in which I drank it. Was it at a pub? Was it served cold? Was I out with a group of friends or on a dinner date? All of these factors contribute to how much I like or dislike a beer and all of them are legitimate. Beer is a culture, not an end in itself. Without the culture around it, it wouldn't nearly mean as much to us, nor taste as good.

    So I guess what I'm saying is that I would actually enjoy the subjective descriptions that would be a part of a beer genome project.

    ReplyDelete
  9. Jack, thanks.

    Shannon: now, you're not biased, are you?

    Shawn, the virtue of Pandora comes in two flavors: 1) consistency; all the genomics are based on a specific set of criteria; 2) minimum level of competence. Let's leave aside the deficiencies of random raters; we can even see deficiencies when we think of our own history of tasting beer. Not everyone knows what to look for or how to find it. Musicologists were harnessed for Pandora because they had formal training; that's what I think would be useful in a Beer Genome Project. (And probably I wouldn't make the cut.)

    Kyle, more or less see above. Breweries generally write in florid prose to promote their beers, not necessarily to accurately describe it. (Not that YOU'D do that--I wouldn't believe it.) Consistency is key, and also objectivity.

    Bacchation, I have no knowledge/opinion about this, but probably Shannon will be happy to weigh in.

    Brian, good man.

    ReplyDelete
  10. Jeff, musicologists were hired for Pandora because they are probably the only people who can identify syncopated drums and pull-offs with the A Dorian mode. But more than professional tasters should be able to identify straw colored body and floral hop aroma. Maybe some people can't pick up a slight melon finish, but crisp mouth-feel shouldn't be too difficult for a beer-geekish person who is not a professional taster. What I'm saying is that it's MUCH easier to get a lot of people who are only pretty good at beer tasting as compared to some people who are experts at it. Will the results be as good as professional tasters? Probably not, but isn't getting 80% as good better than not getting anything at all because you can't get buy-in from professional tasters?

    ReplyDelete
  11. I love the concept... But it's missing a certain, well, immediacy. You turn on Pandora and you have music.

    You figure out how to put a tap on the front of my PC that actually *pours* me the beer it recommends, and you have your get rich quick scheme!

    ReplyDelete
  12. Jeff, what you're talking about is similar to how Dr. Tom Shellhammer and the folks at OSU Fermentation Sciences taste beer. They do it blind and use very extensive tasting sheets were every flavor attribute is rated from 0 (not present) to 15 (overwhelming). They taste as a panel and then merge those scores. By using predefined terms (and also having a way to isloate certain attributes so every one knows what that flavor means), they can get a pretty decent composite. For example, they have a term called grassy. They also have a sample bottle of an aroma that they've isolated as "grassy." So when they tatse, they taste for "grassy" as they've defined it - not as each taster defines it in their own mind.

    One of the main uses of this methodology is that they can take two samples from a brewery (perhaps that show an ingredient change or production change) and then try to quantify the change. I believe their panels are normally 8-12 people, but don't quote me on that.

    Having done some classes with them, I will tell you, they are very, very critical tasters. They're not "judging" in the sense that BJCP folks do - they're really trying to quantify the intensity or absence of different flavors and aromas in a beer. You might want to speak to him to find out more. He's an incredibly nice guys.

    ReplyDelete
  13. re: bacchation said... 'Pintley also seems to quietly pay for its self by suggesting that I might want to try every single beer that Sam Adams' . . .

    There may be an non-conspiratorial explanation.

    Don't know.

    But, in the backwater of SW Fla where I overwinter, Samuel Adams is the only craft brewery readily available with more than [say] 3 different styles.

    I'm glad they're available. Some Sam Adams, some Victory, some Terrapin, some Full Sail, some Redhook, some Kona, some Shipyard, and, only recently, Cigar City.

    It may be sunny and warm; but, it's not Beervana.

    ReplyDelete