I’ve spent a fair amount of time looking at predicative analytics. Predicative analytics is made of software, data, and a model that should forecast, or predict, something. Consumer goods and retail companies use predicative analytics to try and determine what someone will want to buy. Media properties use predictive analytics to take a good guess at what you would like enough to consume. In other words, predictive analytics tries to determine what you want. This is tricky and sophisticated stuff. It requires lots of data about individuals and people like them, an understanding of the patterns that indicate taste, and software capable of processing all of that data incredibly fast. Which brings me to Spotify.
Full disclosure: I love Spotify. I love it enough that I’m willing to pay for it. On Spotify, I can find almost – almost – any kind of music I might want. And I like a lot of music. My taste is, to say the least, eclectic. Progressive rock from my early youth, punk and new wave from my teen and college years, electronic dance music (which everyone knows is the successor to 80’s techno), 90’s alternative, modern indie, blues, early music (as in 1500’s), reggae, rockabilly and so on. When it comes to music, I’m certain to give predictive analytics engines a fit. I recognize this upfront.
Spotify has been doing a lot with analytics of late. Recently they released something called Taste Rewind. Their software asks a user to pick a few bands from a group they display and then generates a playlist from each decade from the 1920’s onward. That’s a pretty straightforward analytics and big data task. That’s not to say it’s easy to do but at the heart of it, Taste Rewind is comparing the characteristics of the artists that you choose with other artists in their database for similar matches.
Last week they added something new to their service. Called Discover Weekly, Spotify puts together a weekly custom “mixtape of fresh music”. With Discover Weekly, Spotify is trying to predict music that I want to listen to. Based on my listening history, and comparing it to others with similar taste, they put together a playlist that should – and I mean should – give me music that I want and have not heard.
Alas, it’s close but not there yet. For the most part, the music is serves up is music I like a lot. That’s good. I don’t like all of their picks but I think that’s impossible. No forecast, whether human or software generated, is going to be 100% accurate. The bad is that little of it is particularly new to me. Of the 30 or so songs in a Discover Weekly playlist, on average I’ve heard probably 28. The worse is that, while it’s all good music, it’s not necessarily what I want to hear. Just because I like The Clash doesn’t mean I particularly want to hear them now. Instead of Discover Weekly, it should be called “Music that you probably like but might not be in the mood for right now” Weekly. That would be a terrible brand name though…
And this is the heart of the problem with predicative analytics. Humans are so complex that trying to determine what they want is too hard to do. It’s a gamble at best. Discover Weekly probably works for a great many people, especially those with limited taste. For listeners who have wide tastes and really get into their music it’s probably impossible to predict what they want to hear at any one moment. Unfortunately, those are likely to be Spotify’s best customers. They are the customers who really want to discover new and interesting music, and are willing to pay for it, whereas the teenage Beyoncé fans only want to know what’s popular.
Discover Weekly should get better over time as Spotify refines the data model and adds new sources of data. It will never be 100% accurate but hopefully will figure out how to server up really new tunes.
It will be at that moment when they kick Apple Music right to the curb.