Spotify’s Music Recommendation
http://benanne.github.io/2014/08/05/spotify-cnns.html
In class we discussed the importance of recommendation systems as they are essential to many markets and is pivotal to many companies success such as Amazon’s e-Market, Netflix, Spotify, Pandora, and many other platforms. Many of these companies have distinct methods and algorithms to conduct their recommendations, usually which require hundreds upon hundreds of models that are used in conjunction to operate. Spotify’s main methods include collaborative filtering, content-based filtering, and deep learning.
Collaborative filtering is focused on using historical data from a user to understand their preference and interests and basically builds a “profile” for a user. For example, if two users listen to largely the same set of songs, their tastes are probably similar. Conversely, if two songs are listened to by the same group of users, they probably sound similar. This kind of information can be exploited to make recommendations. Pure collaborative filtering focuses on its abstraction so that it can be applied to any group or market, whether that be books, music, food, movies, clothing etc. However, this is difficult because popular connections are driven by data, and those yield boring and predictive results. The opposite is not handled as well, as it is difficult for example, to recommend new artists and songs since there has been no historical data on them.
Content-based recommendation revolves around the data about the item itself, in this case the way the song sounds, what genre it is, and thus building up recommendations in a fashion that songs which are similar will share attributes that the user may like, or can be grouped together as such. But even with data on the instruments used, genre and mood it makes it very difficult to assess whether the song will still sound good or not.
Finally, with deep learning, many collaborative filtering models work together analyzing their listeners and songs into “latent Space”. The position of a song in this space encodes all kinds of information that affect listening preferences. If two songs are close together in this space, they are probably similar. If a song is close to a user, it is probably a good recommendation for that user.
A combination of these models work well in practice as the latter two are useful in understanding a new user, and the first can be used well after data has been collected on them.