How Spotify Recommends New Songs
Doug Schwartz
9/11/2020
Article Information
- Title: Spotify’s Discover Weekly explained — Breaking from your music bubble or, maybe not?
- Author: Valerio Velardo
- Date: Feb 11, 2019
Check out the article here.
Summary of Article
This article focuses on the three recommendation models that Spotify has combined to in order to generate its Discover Weekly playlist. For those who aren’t familiar, every Monday, Discover Weekly offers 200 million Spotify users a personalized playlist of thirty songs they’ve never heard before! Though the three recommendation models employed are used by other industry players, Spotify has combined them uniquely into a uniquely powerful music recommendation engine. We’ll take a look at each of these:
- Collaborative Filtering consists of collecting and analyzing users’ behaviors
- Natural Language Processing (NLP) looks at the descriptions of songs and artists (a.k.a. Content-based filtering)
- Convolutional Neural Networks (CNN) are extracted from the raw audio through machine learning (a.k.a. Audio Features)
The 3 Recommendation Models
The Spotify music recommendation framework
Collaborative Filtering
This involves comparing a user’s behavioral trends with those of other users. Content streaming platform Netflix similarly adopts collaborative filtering to power their recommendation models, using viewers’ star-based movie ratings to create recommendations for other similar users. While Spotify doesn’t incorporate a rating system for songs, they do use implicit feedback – like the number of times a user has played a particular song, saved a song to their lists, or clicked on the artist’s page upon listening to the song – to provide relevant recommendations for other users that have been deemed similar.
Source: How AI helps Spotify win in the music streaming world
Natural Language Processing (NLP)
NLP analyses human speech via text. Spotify’s AI scans a track’s metadata, as well as blog posts and discussions about specific musicians, and news articles about songs or artists on the internet. It looks at what people are saying about certain artists or songs and the language being used, and also which other artists and songs are being discussed alongside, if at all, and identifies descriptive terms, noun phrases and other texts associated with those songs or artists.
Source: How AI helps Spotify win in the music streaming world
Example of NLP Data - terms and weighted scores
Convolutional Neural Networks (CNN)
Spotify uses convolutional neural networks to extract musical features directly from raw audio. Interestingly, convolutional networks have mainly been used with visual data. As a result, data scientists have successfully applied them to image detection. This is achieved by feeding a dataset of images to the network, pixel by pixel, to train the model. Once trained, the algorithm is capable of classifying different objects that appear in images that are new to the network. In the case of Spotify, the network has been modified to accept audio data as the input instead of pixels.
The CNN model is most popularly used for facial recognition, and Spotify has configured the same model for audio files. Each song is converted into a raw audio file as a waveform. These waveforms are processed by the CNN and is assigned key parameters such as beats per minute, loudness, major/minor key and so on. Spotify then tries to match similar songs that have the same parameters as the songs their listeners like listening to.
Convolutional Neural Networks are used to hone the recommendation system and to increase accuracy because less-popular songs might be neglected by the other models. The CNN model ensures that obscure and new songs are considered.
Sources: How Spotify Uses Machine Learning Models to Recommend You The Music You Like; and Spotify’s Discover Weekly explained — Breaking from your music bubble or, maybe not?
Plot of the output of the network for 30 seconds of ‘Around the World’ by Daft Punk.
Popularity of Discover Weekly
In the five years since its launch, listeners have also streamed endless hours of the Discover Weekly playlist—over 2.3 billion hours between July 2015 and June 25, 2020 For the numerically inclined, that’s more than:
| Days | Weeks | Years |
|---|---|---|
| 97300000 | 13900000 | 266500 |
Source: Spotify Users Have Spent Over 2.3 Billion Hours Streaming Discover Weekly Playlists Since 2015
What Do I Think?
The article gives an excellent overview of three different types of popular recommendation models. However, I thought the end of the article provided the most insightful portion in which the author discusses how the combination of the three models was essential to Spotify’s success. Different users want different things out of their recommendations. If all the songs are too similar, users will complain the new songs are redundant; if songs are too different, users complain it’s not identifying their tastes. While the in-depth analysis of the recommendation models highlights the science of data science, finding the right balance of similarity and difference underscores the art of data science.
