- The paper presents a novel system that integrates generative AI and collaborative filtering to convert text queries into personalized playlists.
- It employs a three-stage methodology comprising tag extraction, personalized track retrieval, and LLM-based refinement of playlists.
- Deployment results on Deezer show a 45% follow-up listening rate, outperforming traditional manual playlist curation.
An Evaluation of Text2Playlist: Advancing Personalized Playlist Generation on Deezer
The paper, "Text2Playlist: Generating Personalized Playlists from Text on Deezer," proposes a novel system that enhances user experience on the Deezer platform through the creation of personalized music playlists. This system, named Text2Playlist, capitalizes on generative AI, music information retrieval, and recommendation systems, making it a progressive entrant into the domain of music streaming services.
Introduction and Motivation
Deezer, an international music streaming service, has traditionally depended on search functionalities to help users navigate its extensive catalog. These search systems are typically optimized for narrow intent queries, characterized by a precise, navigational goal. However, 20% of user queries on such platforms are broad intent queries, aimed at satisfying informational goals rather than targeting specific tracks or albums. Recognizing this gap, this paper introduces Text2Playlist as a solution for handling these broad queries, facilitating seamless user experience by translating text descriptions into playlists, thus enriching personal music discovery without manual curation.
Technical Framework
Text2Playlist operates on three principal components:
- Tags Extraction and Tracks Retrieval: Using LLMs, explicit and implicit tags from the user's text query are extracted. The extracted tags, such as moods or time periods, are matched with metadata from Deezer's music catalog to retrieve relevant tracks.
- Collaborative Filtering for Personalization: Employing Collaborative Filtering (CF) techniques, specifically latent models, user preferences are leveraged to personalize the playlist. The similarity between user profiles and tracks is computed, resulting in prioritized track suggestions.
- Tracklist Refinement Using LLMs: Inspired by RAG and two-stage recommendation strategies, a final refinement step uses LLMs to enhance and curate the playlist. This step ensures diversity and quality, optimizing the playlist against user preferences.
Deployment and Results
Following successful internal testing phases, Text2Playlist was incrementally deployed, initially to 5% and later to 20% of Deezer's premium users. The deployment architecture involves Kubernetes clusters for the system's Python codebase, Elasticsearch for tag-search operations, and Faiss library integrations for similarity computations. Empirical data indicates that generated playlists achieved a 45% follow-up listening rate, surpassing the 27% for manually created playlists. This suggests a substantial user engagement level and enhanced satisfaction.
Implications and Future Work
Practically, Text2Playlist represents a step forward in automating playlist curation, reflecting user mood and contextual preferences through textual descriptions. Theoretically, it contributes to the refinement of LLM applications in recommendation systems and explores innovative intersections of AI and music information retrieval.
Future directions may include expanding the diversity of the extracted tags through advanced LLM capabilities or incorporating lyrics and additional metadata for richer track representation. Furthermore, there exists potential to transform Text2Playlist into a conversational assistant, enriching the iterative user interaction model and adapting to dynamic user inputs.
In conclusion, Text2Playlist substantiates the practical benefits of integrating generative AI with music streaming services, providing a tailored approach to enhancing user engagement through the natural articulation of music preferences.
