Spotify’s musical network
Networks are implemented in a vast number of fields. Friendship networks on Facebook, communication networks in Version, and even political networks in campaigns. It is almost universal in any large technological service. So it is not a surprise to know that Spotify uses its user network to recommend music.
The ability to graph relationships between elements proves to be very useful information when trying to predict trends and patterns. Much like how Facebook uses its large user database to recommend friends, Spotify collects information from its users to recommend music for other users who have similar tastes. This type of recommendation algorithm is called “Collaborative filtering” and in essence, what we have is a network where the songs are the nodes and the edges are users who have listened to both song A and song B. The more people that listen to both A and B the stronger the tie is between song A and B, which gives Spotify the advantage in provided good recommendations.
Good recommendations using this prediction path tie closely to the Triadic Closure principle. Since a user listens to both song A and B, it is highly likely that there should be an edge between song A and B. If more users display the same trend then this edge between A and B would become stronger and stronger, thus making the recommendation a good recommendation. It is a simple network that draws upon the basics that we have learned in class.
However, this network might be too simple as the article points out. Just drawing from the principle of Triadic Closure and “historical usage data” can cause recommendation issues like being “content-agnostic,” “heterogeneity of content with similar usage patterns,” and “new and unpopular songs cannot be recommended.” Much like how when we added information to the network to contain strong and weak bonds, we were able to produce a better prediction principle, the Strong Triadic Closure principle, Spotify implemented a new network which incorporates more information based on the song’s contents rather than user preferences to better recommend songs. With more factors to consider between each node, the prediction patterns become more robust and produces better recommendation.
While the article goes more into depth about how it implements a recommendation algorithm that predicts “listening preferences with deep learning,” the large picture takeaway from this is how the accuracy of prediction rises as more information are considered in each relationship. The hard part of prediction is knowing how to incorporate as much information as possible in the network, the easy part is using those relationships and the principle we have learned in class to provide good recommendations.
Source: http://benanne.github.io/2014/08/05/spotify-cnns.html