Spotify shares its secret sauce: 3 layers of AI-driven algorithms
Spotify’s AI-driven algorithms have drawn praise for their accuracy and concern over their potential for bias. With 500 million active monthly users, how its algorithms work has massive implications for music and global culture.
Now Spotify has offered a detailed look at how it uses machine learning to create each user’s experience on the platform, including what gets played and why is a new WSJ video (watch below).
Spotify’s secret sauce is how it layers one algorithm on top of another to discover uncannily accurate personalized music recommendations.
Layer 1: Collaborative Filtering
Collaborative Filtering looks at when a song is played or playlisted next to another song. “Think of it like building a map.”
Layer 2: Content-Based Filtering
Content-Based Filtering gathers metadata and executes a basic audio analysis to determine things like danceability and loudness to describe the sonic characteristics of a track., It also analyzes the lyrics and adjectives used to describe the track in online articles and logs.
One problem with content-based filtering is that it can reinforce existing musical and cultural biases in a kind of feedback loop.
Another is the “cold start problem” – the lack of data for new artists. Here, Spotify’s human curators step in to find the best new music.
Layer 3: Reinforcement Learning
Reinforcement Learning is the newest layer and growing in importance. It enables Spotify’s entire recommendation engine to learn and improve based on feedback.
Tomorrow: Spotify’s Algorithm from an Artist’s Perspective
This 7-minute video offers a solid crash course in Spotify’s algorithm.
- 0:00 Spotify is known for personalized playlists made with its recommendation algorithm
- 0:37 The evolution of Spotify’s recommendation algorithm
- 1:42 Spotify uses user data to create a map of songs and artists
- 2:59 How Spotify’s content-based filtering tech works
- 4:09 The dangers of algorithmic bias
- 6:19 Spotify’s future plans in AI