Spotify Recommendations
Social networks like Facebook deal with an astronomical number of users and the problem of giving friend recommendations based on their local network. A slightly more conceptually challenging problem is that of Spotify and their music recommendation services. Dead-set on making the way we listen to music a social experience, Spotify supplies everything from related artists, a discover page and weekly playlist, and radio based off anything from a song, artist, or playlist. But how does it go about doing this.
Well, the idea behind it is very similar to how we discussed a possible friend recommendation generating algorithm for Facebook. If A and B have a close tie and B and C have a close tie, linking A and C together would satisfy Strong Triadic Closure. Essentially, Spotify wants to do this for songs (see slide 5), but finding similarities between people, songs, artists, etc. can prove to be much harder to systematically define compared to seeing how often two people like each other’s posts and such. The way Spotify faces this problem is by doing some fairly complicated matrix multiplication. It first makes a matrix of every user by every musical item on Spotify and fills it with how many times the user listened to the item (slide 19). It then tries to create a vector for a user or item (with something like 40-200 dimensions) which more or less sees the user or item’s similarity and differences to those of the matrix and then gives similar objects similar vectors. With these vectors, you can simple give a value to similarity by taking the cosine of their vectors (example on slide 30). This then, in a way, makes a network where every user and item has a connection of varying similarities that can be computed. If there is a strong similarity from you to a song you haven’t heard (or heard in a while), Spotify will send it your way.
Source: http://www.slideshare.net/erikbern/mlhadoop-nyc-predictive-analytics-2?related=3