Researchers have introduced Sona, a novel generative recommender system designed for Yandex Music. This single-model architecture unifies candidate generation and ranking by using a shared user representation, transforming engagement events into hidden states for both an autoregressive decoder and a ranking module. In an online A/B test, Sona outperformed the existing multi-stage cascade, leading to significant improvements in key engagement metrics such as active users, total listening time, and likes. AI
IMPACT This unified model architecture could influence future recommender systems by simplifying complex cascades and improving efficiency.
RANK_REASON The item describes a new model and its evaluation in a technical report, fitting the research category. [lever_c_demoted from research: ic=1 ai=0.7]
Read on arXiv cs.IR (Information Retrieval) →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
