Skip to main content

Spotify’s Predictive Personalization

In the eight years since it’s founding in 2006, Spotify has grown rapidly to become the 658th most popular website globally [1] and now services users around the world. It was one of the first music providers to allow its users to play songs of their choice, entirely for free. This novel approach secured Spotify a loyal fanbase, and it continues to pioneer in the field of streaming music services.

The article Spotify Knows Me Better Than I Know Myself [2] highlights one of Spotify’s most promising future services: predictive personalization of playlists. In the article, Walt Hickey writes of his experience beta testing the new tool. He received not only song suggestions provided by the service, but also explanations of how the suggestions were made, provided by the lead developer. Hickey was surprised to learn that his listening habits are best modeled by two major clusters, including one he endearingly referred to as “The Shame Cluster.” However, after listening to his personalized playlists, Hickey was convinced of both the tool’s accuracy and — more importantly — success.hickey-feature-tasteprofile-table-5

This predictive power stems from how the service characterizes music preferences as a network of songs: songs whose connectivity depends not only on their play counts and genres, but the contexts of their plays. As an example, Hickey learned that though he played a considerable amount of Vitamin String Quartet, it would not necessarily be considered part of his music identity because each song averaged only 11.5 plays. These songs only fit when Hickey is in a certain mood or writing a long post. A generic classical playlist would not appeal to him. These considerations are at the core of the tool’s eventual goal of ideally predicting what songs are appropriate on a Friday night vs. a weekday morning. Each song’s connection to the others depends on the context in which they are played.

Spotify’s tool can easily be related to the concepts of weak ties and local bridges when considering how the algorithm suggests new songs not in the user’s play history. Even if two clusters involve very similar artists or genres, they may only be weakly tied if they are played in significantly different contexts (e.g. studying vs. partying). When suggesting new songs that would help satisfy triadic closure, the algorithm could consider only ties that would help satisfy strong triadic closure. Thus, local bridges between contextually distinct clusters would not provoke a new song suggestion if the bridge is a merely a weak tie. In this manner, the tool could apply the concepts of weak ties and local bridges to successfully avoid suggesting similar songs inappropriate for a given context.



Leave a Reply

Blogging Calendar

September 2014
« Aug   Oct »