Paris, Île-de-France, FranceIntern
Just Hack it!
With 53 millions of tracks and a presence in 180 countries,
Deezer is the most personal music streaming service in the world.
Behind the code and the pixels is our team of 500 music lovers, and we’re building something incredible together. Want in? If you’re looking for an adventure, not just a job, and you fancy seeing ideas come to life in a heartbeat, you’re in the right place.
We dare to challenge the status quo and believe innovation is part of our DNA.
Before streaming services, the problem of finding the right track ordering was referred to by radio music programmers as "creating a clock". The general idea is to alternate Hits and Discoveries, in order to keep users engaged. Moving forward, we can nowadays pick from additional user-specific pools of tracks (recommendations, recent discoveries, favorites, …) and have access to user's listening history and interactions. Can we better combine tracks than using a blind alternating principle?
Optimal ordering is typically NP-hard (eg. Travelling Salesman Problem). When an explicit optimal ordering exists, many methods are able to compute or learn approximations for it. For instance, several works have trained neural networks to decipher sentences from a scrambled list of words [2, 3]. But music is way more implicit, subjective and may depend on the context. While it is hard to define what an optimal ordering of tracks should be, local ordering and transitions do matter .
From this perspective, some approaches have considered computing track features (tempo, key, ...) and handcrafted metrics (novelty, diversity, ...) to greedily reorder tracks to ensure coherency between successive ones while maintaining the overall playlist logic . More recent reinforcement learning works focus on selecting music from the right pool of tracks using bandit algorithms  or generating next tracks on a direct sequence-level .
The intern work will be to extend the existing research in the field of track reordering. We envision the outcome of this project to be either a prototype which could be later transferred in production-ready code or a scientific article submitted to an international conference.
The intern is supervised by research scientists and research engineers from the Deezer R&D team who provide practical and scientific help with the performed task. The intern is nonetheless encouraged to propose solutions and work autonomously. For data experiments, Deezer ensures cutting edge technology and appropriate calculus power.
 Scaling Up Music Playlist Generation - Aucouturier, Pachet (2002) https://ieeexplore.ieee.org/document/1035729
 Order Matters: Sequence to sequence for sets - Vinyals et al. (2015) https://arxiv.org/abs/1511.06391
 Implementation example: https://github.com/Kyubyong/word_ordering (2017)
 An Analysis of Approaches Taken in the ACM RecSys Challenge 2018 for Automatic Music Playlist Continuation - Zamani, Schedl, Lamere, Chen (2018) https://arxiv.org/pdf/1810.01520.pdf Spotify playlist continuation challenge here: https://recsys-challenge.spotify.com/
 Explore, Exploit, and Explain: Personalizing Explainable Recommendations with Bandits - Mc Inerney et al. (2018)
 Automatic, Personalized, and Flexible Playlist Generation using Reinforcement Learning - Shih, Chi (2018)
Master student with a background in Computer Science / Applied Mathematics / Statistics
Knowledge in Sequence-to-sequence problems, Reinforcement Learning
Strong knowledge of applied machine learning and data mining
Good programming skills for data processing and experimentation (preferably python, but we are open to other technologies too)
Life @ Deezer Paris