If you want to demo some technology, you might as well have fun doing it. We found an opportunity for some fun in the form of a great American Whiskey bar named Char No. 4. Based in the Cobble Hill neighborhood of Brooklyn, it’s the creation of former professional sommelier Sean Josephs. With Sean’s help and tasting chops, we built Whiskey Engine on the Axilent platform, a source for recommendations on American whiskey.
Whiskey Engine is extremely simple in its concept. In its catalog is the vast majority of Bourbon, Rye and Kentucky Whiskey available. If you’re having a look at a particular whiskey, it will show you other whiskies that are similar to it, and if you indicate that you like a whiskey by leaving a positive review, it will show you other whiskies that you might also like.
Modeling Whiskey
To understand how whiskies related to each other, we modeled the characteristics of the taste of whiskey itself, thus giving us a foundation on which to make recommendations. To get the taste data, Sean and his team tasted about 84 whiskies from their shelves. Sean, the domain expert, defined the taste of whiskey in eight distinct terms: herbs/spice, flowers, fruit, candy, wood, length of finish, complexity and smoothness. For each of the tasted whiskies, they ranked each of the flavor elements on a scale from 1.0 to 5.0.
With such great structured data available, we decided to implement the relativity of whiskey by using metadata. (In cases where the data structure is indeterminate or inconsistant, strategies based on tagging can be better). We modeled whiskey as a Content Type in Axilent, and applied the tasting data as metadata field values for each whiskey.
However, we suspected that not all taste elements were equal when it comes to determining what people like. Consulting with Sean, we found out that there is indeed a significant inequality in the elements that drive people to enjoy a particular whiskey. Listed below, in order of most to least important, are the taste elements of whiskey, and the saliency values (on a scale of 100) we assigned to each element:
- Smoothness 90
- Complexity 80
- Candy 70
- Length of Finish 60
- Wood 50
- Herbs / Spice 40
- Flowers 40
- Fruit 40
As you can see, smoothness is king. When in comes to people’s affinity for a particular whiskey, it is the most important element of its taste. People who like a whiskey will tend to like it because it’s smooth, or because it’s rough. Based on what they said they liked, the best bet is to find another whiskey with a similar smoothness profile. Also high up on the scale are complexity, candy (“it tastes caramel-like”) and the length of finish. Down at the bottom are elements like herbs/spice, flowers and fruit, which can be far less perceptible for inexperienced palettes.
As implemented on Axilent, if you were looking at a whiskey that had a 4.0 out of 5.0 for smoothness, we should try hardest to find other whiskies with a 4.0 smoothness, before looking for for matching flowers, fruit etc. Smoothness is just more important to most people’s tastes. This is a great example of why we’ve built the concept of user-definable saliency for content type fields into the Axilent platform.
Making the Recommendations
For personalized recommendations, we leveraged the same metadata. In Whiskey Engine, you can leave a review of a whiskey. If you leave a 4 or 5 star review, we interpret that as your endorsement of that whiskey. From there we look for for whiskies that are related by their taste profiles to the whiskies you’ve endorsed. Instant product recommendations.
We defined two content channels, both extremely simple. The Related Whiskey channel (populating the top row under the featured whiskey) is entirely field-related content, using the featured whiskey as the base content. The related fields are the taste elements described above. The second row is populated by Personalized Whiskey, which is (as its name implies) entirely personalized content.
Whiskey Engine is a great example of how a nuanced recommendation product can be built easily on the Axilent platform, and is a lot of fun besides. Now, if you’ll excuse us, it’s time for some Hudson Manhattan Rye. Cheers.
If you enjoyed this article, please consider sharing it!
