A simple Google search will quickly reveal that Portland, OR, (aka “Beervana” or the “Hub for Hops”) ranks top among the best beer cities in America! But if you’re anything like me — someone who really enjoys a great beer or wine, but simply doesn’t have the time or capacity to find what you like most among the plethora of fantastic local options available here — there is finally a solution. Next month, Next Glass will launch their iOS app, which leverages augmented reality functionality (and under-the-hood, patent-pending chemical science and machine learning), to deliver highly-accurate beer and wine recommendations to users on their mobile devices.
Personalized ratings on beer and wine
For this month’s second of three, New Relic FutureTalks PDX Summer Series event, we were lucky enough to have Next Glass’s own COO Trace Smith (a.k.a. the “guy who buys all the beer”) explain how each of these elements (science, machine learning, and augmented reality) benefit and empower beer and wine consumers. Check out the full recording right here:
In a post-event download conversation with his team, Trace emphasized that he wanted talk attendees to walk away with a sense of how science and technology can combine to be a disruptive force in an industry sorely in need of change. An added bonus for all those on hand for this packed event was a chance to see Next Glass in action, and to be the first to receive access to the beta later this month!
Data mining your taste buds
Prior to Trace’s stop in PDX, we asked him a few questions about the science and technology powering their service. We got the down low on everything from their user ratings and personal preferences to the data that drives their recommendation engine:
Can you give us any details on the algorithm you’re using for drink recommendations?
Our liquid chromatograph high-res accurate mass spectrometer (which we affectionately call “Corky”), runs a full scan for 22,000+ compounds on each bottle of beer and wine we test. Each of those compounds is not present in each bottle and not all of the compounds that are present contribute to flavor (taste or olfactory), but we can’t cherry pick the several hundred that do drive enjoyment. The high-res mass spec allows us to measure both presence and abundance of each compound, which helps — binary data (presence of a compound or lack thereof) isn’t enough to provide accurate recommendations.
We have a ton of data.
There are 30+ algorithms we run the data through and essentially layer machine learning on top of machine learning to determine which data points optimize recommendations for our users. (Least variance and quickest delivery are the things we’re looking for).
How many tested samples does it take before you can recommend a product that we’ll like (maybe 99% of the time)?
If you give us a single beer or wine you love, we can recommend you something else you’re certain to love. Once you rate a few things you do and don’t like (say, three you do and one or two you don’t), we can begin to open up our whole catalog and tell you how likely you are to like anything we’ve tested. There’s not much up-front friction for the user. One really neat thing we realized early on is that we can tell you wines of a certain varietal (or beers of a certain type) that you’ll enjoy even if you’ve not provided us a rating for a wine of that varietal. So, if we know a few Pinot Noirs and Chardonnays you do and don’t like, we can recommend some Cabernet Sauvignons and Pinot Grigios you’ll enjoy, and also steer you clear of some you won’t.
Are there a specific set of data points that give you the best indication of personal preference?
Once a user starts to provide us ratings, we can see chemically what they’re liking or not liking in wine and/or beer. The data points are totally idiosyncratic on the user level – meaning the compounds that others enjoy or dislike may not be that important to you. That said, there are several hundred that our machine learning efforts have told us matter across a broad swath of the population.
Are you able to correlate “hint of cherry and oak” with the molecular data?
We haven’t tried — what we’ve done is eliminate the need for the frou-frou adjectives that only serve to confuse your average wine drinker (and, to a lesser extent, beer drinker). I’m not saying there’s no merit in those incredible descriptions, but I’m definitely not able to discern a “hint of eucalyptus on the nose and smattering of black fruits on the palate,” and I don’t think most wine drinkers are able to either.
I’ve certainly run across extraordinary people who have incredibly discerning palates, but most of us can tell whether or not we’ve enjoyed a beverage and not much beyond that. We think that’s all anyone should need to be able to do to discover new great beer and wine.