The market for streaming audio is hotly disputed. To differentiate themselves, companies are betting on the branch resources enhancement. Take for example Spotify, among other things, stands for the suggestions of playlists. And this functionality might be better: the company wants to use artificial intelligence to make recommendations to approach more of the tastes of each user.
The project was described by Sander Dieleman, doctoral student at the University of Ghent, Belgium. The student doing a summer internship in the Spotify office in New York and published a very enlightening post about your work there.
Dieleman explains that the main goal of the project is to make lesser-known songs reach users. There are good quality songs in the collection of Spotify, but not implemented due to the very fact of their authors / performers are little known.
If many uses streaming services precisely to meet artists and musical styles, there’s something wrong with this approach. The Dieleman own points which is part of the problem: the use of “collaborative filtering” in recommendation systems.
This technique analyzes the history of executions of each user to determine their preferences. If, for example, two of them have a large number of listening to music in common, the system considers that they share similar tastes. Thus, if one has heard a song the other did not hear, the system can recommend it to the first.
Most often, this method results in good suggestions, but there is a weak point: if users with similar preferences hear only the best known songs end up being slightly less popular tracks recommended.
In late May, when Spotify officially debuted in Brazil , the director for Latin America of the company Gustavo Diament explained that there are teams dedicated exclusively to the creation of playlists. This point alone shows how important are the recommendations in the service.
Suggestions to improve the team which is part Sander Dieleman is focusing on “Deep Learning” (Deep Learning), algorithmic technique that analyzes levels in a series of data to learn to distinguish patterns. Google, Microsoft and Netflix are among the companies that have bet on this method.
More precisely, the system did Dieleman analyze samples 30 seconds of the 500 000 most popular songs in Spotify. Thus, the service can recognize similar patterns in less known bands and then recommend them.
The system is under development and still needs a lot of training, but preliminary results are encouraging. According to Dieleman, filters already can generate playlists based on various characteristics, such as voice patterns, chords and distortion.
Dieleman claims to have found yet filters generated through the technique that succeeded in classifying rhythms like reggae, salsa, punk, and Turkish pop (!!!), all without errors.
It is unclear when the system will go into full operation. When that day comes, Sander Dieleman hopes that Spotify can not only hit on the recommendation of lesser-known songs as well as songs that prevent little or nothing consistent to the user’s playlists in between. We also hope 🙂