Sites Inria

Version française

Collaboration

20/06/2014

Deezer: listening to you!

© Inria / Jérémie Mary

Deezer, the French leader in Web-based music streaming, is attempting to win over the American market. With Inria, the company is developing a music recommendation system to attract new listeners.

To stand out in the American online music market, Deezer is turning to new services. Among these innovations is a music recommendation system. The Sequel* research team at Inria Lille–Nord Europe (working with the Centrale Lille engineering school and Lille 3 University*) has been set the task of developing this new feature. "It's all about suggesting users listen to tracks selected to suit their tastes as much as possible", explains Jérémie Mary, senior lecturer at Lille University and member of the Sequel team. "The challenge lies in striking a balance between exploration and exploitation—enabling listeners to discover new tracks they might like so we can better define their profiles, while not making that seem like a discomforting intrusion."

The solution Inria offers is based on the listening data Deezer provides. Cross-checking is carried out based on tags linked to each track and containing various details (group name, genre of music, etc.). Listening time for a given track is also factored in.

"On Deezer, around 30% of listening time for a given track is 30 seconds and 50% of tracks aren't played to the end. The skip option for navigating around a playlist partly explains this," Mary explains. "What's more, over a million albums are available on Deezer, but only 20% have significant audiences. The possibilities for recommending music are therefore huge. It's up to us to adjust our algorithms to detect the best tracks, without getting lost in this labyrinth of data."

Towards full-scale deployment

The system is currently being tested by Deezer developers and still has a big hurdle to get over: full implementation. "Our solution currently has the particularity of being totally online. Systems working online have the advantage of instantly taking into account users' tastes," says Mary. "Traditionally, when a new album is released, the marketing services promote it. In our approach, we use so-called 'bandit' algorithms to statistically determine how large a given marketing campaign should be and which users it should target. This approach allows us to resolve the 'cold start' problems of recommendation systems. The price to pay is increased consumption of resources. With Romaric Gaudel, we're considering ways of processing the data stream in real time while dynamically adapting to available resources and the standard flow through user aggregation mechanisms. One thing is certain: Whatever solution is chosen, the work carried out will benefit all those involved. Deezer, of course, has used our experience to formalise their ideas, and Inria has been able to finance its research on algorithms."

*At UMR 8146 CNRS-Centrale Lille-Lille1, LAGIS and UMR 8022 CNRS-Lille1-Lille 3-Inria, LIFL.

Keywords: Research project Attractiveness Sequel team Algorithms for collaborative editing Music

Top