We are a leading company in the music streaming industry, with one of the largest catalogues on the market (over 120 million HiFi music tracks) and more than 9 million worldwide subscribers. To help access this vast amount of tracks, recommender systems are a key asset to retain and attract users. Suggesting relevant personalized songs, artists, albums, or playlists helps users actively explore the vast and mostly unknown musical landscape.

At Deezer, the recommendation team is in charge of building machine learning models and algorithms like the Flow to achieve such a goal. For this purpose, transformers have recently emerged as a competitive approach for sequence modeling and generation in various sequential recommendation problems.

Transformers were successfully trained and deployed to perform automatic playlist continuation and considered for other features. They revealed promising performances.

However, they can also exhibit some drawbacks like popularity biases. Consequently, they can impair accuracy and fairness in the recommendation algorithms.

The objective of this internship will be to investigate one of the critical aspects related to the training of Transformers for music recommendation at scale on Deezer: negative sampling.  The way we perform negative sampling during training is crucial. It is known to significantly impact the final performance of the Transformer, as well as its tendency to recommend popular tracks. 

The intern will be supervised by a Senior Data Scientist from the Recommendation team, and will also directly work with other engineers and research scientists from Deezer. They will provide scientific and technical guidance throughout the internship. For experiments, Deezer ensures access to internal data, cutting-edge technology, and appropriate calculus power. Several previous interns from our team have tested their algorithms in production, and/or have published results from their work as scientific articles.

• Master / PhD student with with a strong background in machine learning
• Good programming skills and knowledge of machine learning tools (such as scientific Python, PyTorch, Keras, etc)
• Creativity and autonomy
• Good data processing and analysis skills
• Knowledge and experience in recommender systems would be a plus
• Knowledge and experience in Transformer architectures would be a plus