Give me a song

I listen to music constantly. It is a necessity by now to have music in order for me to do any work. I feel slow and distracted with out it. But I have one problem: with all the music I listen to, my regular tunes can get old pretty quickly. While I love looking for and finding new music, it takes a while to do, and it can often distract from more pressing work at hand. Luckily, Spotify is there to save me.

Spotify https://www.spotify.com/us/ is a subscription music service where you can play music on demand. It will also make playlists for you of personalized recommended songs. These recommended playlists use predictive algorithms and machine learning to aggregate more music that you will (hopefully) like.

It uses three main recommendations models:

1. Collaborative Filtering
2. Natural Language Processing
3. Audio Models

Collaborative Filtering was popularized by Netflix. Users could rate the movies they saw on a 5 star scale. Netflix would then take this rating and recommend more movies based on the user’s rating. The cool part about Spotify is that they add a twist to this: since they don’t have a traditional rating system, they had to come up with other ways to judge how much you liked the song. Mostly, they use how many times you streamed certain songs, but they also take into account other data like how long you listen to the song as well as where you went from that song. This model is extremely popular and is usually the building blocks for most modern recommendation systems.

Natural Language Processing is when the computer reads literal english and “understands” it and its sentiment. Spotify utilizes natural language processing by scraping the internet for articles and blogs about music to see what music or artists are hot right now, helping it to suggest them to you. By tracking terms and sentiment, Spotify can keep a finger on the pulse of the online music work and keep you up to date with the new, hot music.

Raw Audio Models take in the actual audio from songs and analyses that. This targets music that could be passed by in the other two models. If a song has only a few plays and isn’t getting much love on the internet, it very well may fall through the first two models, even if it would be perfect for you. In order to catch those songs, Spotify uses neural networks to analyze the raw audio data of these songs. It does not discriminate between new or old and popular or unpopular, so it can catch the songs that fall through the cracks above. The model spits out some characteristics of the song like time signature, key, type, tempo, etc. Spotify can then use this categorize the songs and recommend you songs in categories that you like.

The world of music has an insane amount of data. Even if you know what you like, there is no way that you can filter through all of the music yourself. Spotify is on your side, using analytics, to help you find your next favorite song.