Yandex announced that it consolidated a multistage recommendation system into Sona, combining an encoder, a generative decoder and ranking. It represents content with learned semantic identifiers to reduce the complexity of manually engineered features and multistage candidate processing. The company reported a 6.3% increase in listening time and a 4.5% rise in activity during a seven-day A/B test of music recommendations on smart speakers. These are company-reported results for a specific product, not evidence that the same improvements apply to other content or long-term use.
Back to Tech
Yandex Sona unifies recommendation candidate generation and ranking in one model
Reducing recommendation stages is noteworthy, but a short trial on one service cannot establish overall recommendation quality.