Recommendations on the Internet suck
I maintain a list of over 100 TV series and movies to watch spending countless hours reading reviews because the best Netflix can do is recommend me El Chapo since I liked Narcos. Amazon recommends me a primer on economics because, well, I just bought a primer on economics. The recommendations that YouTube makes has made me develop a “banner blindness” to them. Travel sites recommend me hotels that I would definitely stay in when I turn 60. Zomato is best used to get a list of places that I can then ask my friends about. And don’t even get me started on the recommendations on Apple Music.
This in a day and age when the world is collectively orgasming over machine learning and recommendation algorithms, and companies are putting up million-dollar prizes to improve them.
Recommendations have to go much deeper than what I mention my interests are, the pages I like or the links I click. When I say I like mountains, it means I like to stay in places with a great view and from where I can easily walk around town but all other things can be basic. When I say I like sci fi novels, it means I like the ones with a great story at their core, rich in detail but with simple language. When I say I like continental food, it means I like continental food in places with a fun vibe, not the uptight fine dining ones.
It also goes much deeper than what my friend graph might like, which is another way a lot of services tackle it. I might go to one friend for mountain vacation recommendations, to another for beach-based vacations, to my sister for the best clubs in town, to a colleague for sci-fi book recommendations, but recommendations based on the friend graph is most likely to bring me the lowest common denominator from all of them.
Solving it is no easy task but I think the upside is high enough to make it a problem worth solving. Bringing in more nuanced data definitely helps make better recommendations, which a lot of services are working on (Netflix doesn’t quite seem to agree though).
But an overlooked part of the problem is understanding consumers and their tastes at a much deeper level. Their beliefs, values, their sense of aesthetics, their idea of good. Hunch tried solving this back in the day by building a taste graph for every user that they could then license to any service wanting to make better recommendations but it didn’t quite work. Studies show that your music preferences tell a lot about who you are and your tastes, so maybe that’s another angle to look at. Maybe the solution is a combination of these and more but I do believe that understanding consumers and their tastes better is the key to making better recommendations.
