Spotify's Recommendation System

Spotify employs the BaRT system - Bandits for Recommendations as Treatments.

The sole purpose of BaRT is to provide users with music that it thinks you will like, and something that you haven't streamed before. Simply put, it tries to keep you listening.

It is a multi-armed system that classifies and ranks tracks as playlists first, and then as shelfs.

BaRT's ranking of cards and shelves make its way to you as your home screen recommendations.

It does this by playing and suggesting songs it knows the user is familiar with, while dropping in some fresh tracks it thinks they might like, but most importantly, haven’t heard before.

This makes sure the listener is engaged, but things never get boring.

BART decides what to suggest to listeners using three main functions:

• Collaborative Filtering Compares new songs to a listener’s current habits to decide what will suit their tastes.

• Natural Language Processing Analyses the language, lyrics and content of a song, as well as searches for mentions of the song across the web - say, in blogs, music articles, news, etc.

• Audio Analysis Detects the “vibe” or “mood” of a song’s audio and decides whether it’s upbeat, chill, heavy, minimal, instrumental etc. Breaks down the audio into music characteristics like key, tempo, etc. to determine similarities between songs.