Recommending Music on Spotify with Deep Learning


(T) Do you enjoy listening to Bach’s Brandenburg Concertos? If you do, you are likely to appreciate Vivaldi’s violin concertos as well. Musicians learn from each other. And, musicians influence each other. So striking similarities can be found between their music. Some of Bach and Vivaldi’s concertos convey the same musicality and leverage the same techniques because of the influence that Vivaldi had on the music of Bach.

Traditional recommender systems provide recommendations through collaborative or content-based filtering or both. Collaborative filtering approaches build a model from a user’s past behavior. Content-based filtering approaches utilize a series of discrete characteristics of an item in order to recommend additional items with similar properties.

There are many features of a song that can be leveraged for content-based filtering: lyrics, artist, album, reviews, date of creation…and the audio signal itself, which is what Sander Dieleman, a Ph. D. student and a Spotify’s intern, is proposing for content-based music recommendations.

Sander is suggesting to Spotify’s users to hear new songs by recommending to them other songs that share some of the same characteristics of the audio signals of the songs they already appreciate. The main goal of that approach is to help listeners discovering in particular new and relatively unheard music.

To that end, Sander is analyzing the audio signals using deep learning algorithms implementing two convolutional nets (convolutional neural nets are widely used for image recognition).

Sander had published the whole fruit of his research on his Google+ Page/GitHub:

And that is a good reading to do :).

Wikipedia: Convolutional neural network.

Note: The picture above is the Spotify apps on my smartphone.

Copyright © 2005-2014 by Serge-Paul Carrasco. All rights reserved.
Contact Us: asvinsider at gmail dot com.

Categories: Deep Learning