Current recommendation systems rely heavily on latent collaborative filtering (CF) models, which analyze usage data to generate low-dimensional embeddings for users and tracks. However, while effective, CF-based models have inherent limitations such as limited coverage of the catalog and slow integration of new releases in the recommendation engine. This internship will address these challenges by exploring multimodal machine learning approaches. By integrating diverse data sources - such as CF, text and audio embeddings - these methods aim to create a comprehensive multimodal embedding space for tracks, leveraged for downstream recommendation / retrieval tasks.

• Conduct an in-depth review of state-of-the-art multimodal methods.

• Design and implement multimodal models for recommendation and/or retrieval.

• Apply these models to real-world Deezer datasets and benchmark their performance.

• Optionally, contribute to publications and / or conduct A/B testing of selected methods in a production environment.

• Master / PhD student with a background in Computer Science / Applied Mathematics / Statistics.

• Strong knowledge of music analysis, natural language processing, applied machine learning and data mining

• Good programming skills for data processing and experimentation (preferred python)

• Creativity and autonomy