The Spotify algorithm analyses audio directly to understand mood

Spotify's recommendation system doesn't just track listening history — it analyses audio features of every track directly: tempo, key, danceability, valence (musical positivity), acousticness, speechiness, and energy. These numerical features are derived by audio signal processing. Combined with collaborative filtering (matching you with listeners who share your taste) and natural language processing of music journalism, Spotify's system can recommend music before enough people have listened to generate collaborative filtering data. The 'Discover Weekly' playlist, launched in 2015, is updated every Monday based on the preceding week's listening.

Source: Ciocca, How Does Spotify Know You So Well?, Medium (2017)