Researchers have introduced Sona, a novel single-model generative recommender system designed for Yandex Music. In an online A/B test, Sona successfully replaced a complex, multi-stage production cascade, which included over 15 candidate generators and multiple ranking models. The new system unified candidate generation and ranking around a shared user representation, leading to significant improvements in key engagement metrics, including a 4.53% increase in Active Users and a 6.30% rise in Total Listening Time. AI
IMPACT This research demonstrates the potential for single, jointly trained models to replace complex multi-stage recommendation systems, potentially simplifying infrastructure and improving performance in large-scale applications.
RANK_REASON Publication of a technical report detailing a new model and its performance in an A/B test.
Read on arXiv cs.IR (Information Retrieval) →
- Argus
- Sona
- Yandex Music
- alphaXiv
- arXiv
- CatalyzeX Code Finder for Papers
- Connected Papers
- DagsHub
- Daria Tikhonovich
- Gotit.pub
- Hugging Face
- Litmaps
- My Vibe
- scite Smart Citations
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →