"Yandex" developed a generative model Sona for the recommendation system

The company "Yandex" presented its own generative model of artificial intelligence Sona, designed for searching and ranking content. The development was carried out by the R&D team of the company, as well as engineers of the services "Alice" and "Yandex Music". The new architecture combines several stages of recommendation selection into a single module, eliminating the need to configure dozens of separate algorithms and manually select features. The system itself analyzes the user's actions, after which it forms the final output.

The first service to test the technology was music recommendations in smart speakers with "Alice". According to the company, during A/B testing, with a complete replacement of the previous chain of algorithms, the listening time of recommended tracks increased by 6.3%, the number of likes increased by 11.42%, and the number of requests to repeat the composition increased by almost 18%. The share of active listeners over the week of the experiment showed an increase of 4.5%.

To select audio recordings, the model uses composite codes Semantic ID, taking into account the track's metadata, the first 90 seconds of playback, and the audience's reaction. For the initial training of the algorithm, the company applied a separate "teacher" model, trained on a yearly array of user actions. In working mode, the service receives an updated version of Sona every 10 minutes based on fresh listens and ratings.