How did Spotify get so good at machine learning? Was machine learning important from the start, or did they catch up over time? originally appeared on Quora: the place to gain and share knowledge, empowering people to learn from others and better understand the world.
Answer by Erik Bernhardsson:
Until about 2012 it really was just me working on it, mostly as a side project. I think a couple of design decisions quite early turned out to work really well, like focusing on collaborative filtering and in particular vector models. Initially PLSA, then later this method. As soon as word2vec came out we switched to it immediately and had it in production just a few weeks later.
Around 2013–2015 I built up a team around music recommendations that focused a lot on related artists, radio, and eventually Discover Weekly (which was released after I left).
We acquired Echo Nest which injected a bunch of smart people into the team but I wouldn’t say the technology accelerated the machine-learning focus. The way I look at it, it was mostly a talent acquisition, not technology. Very little of Echo Nest’s tech actually made it into the Spotify product, most of it being work by their audio team.
We also looked into deep learning quite early, with Sander Dieleman interning in 2014 focusing on pure audio based methods. I’m not sure to what extent this is used in production, but his results were quite impressive. In my mind, the best way to get good recommendations is 90% through collaborative filtering then use deep learning models to get the extra 10%.
It’s been over two years since I left, but AFAIK, Discover Weekly is entirely powered by collaborative filtering, in particular, a few extensions to word2vec that the machine learning team in NYC built.