One AI instead of dozens of algorithms: "Yandex" introduces Sona
"Yandex" has introduced a generative artificial intelligence model named Sona to prepare recommendations for users. According to the company, it can replace a system consisting of dozens of algorithms by combining the tasks of finding suitable options and sorting them. During a one-week trial on smart speakers with "Alisa", the listening time of recommended music increased by 6.3 percent.
One AI instead of dozens of algorithms: "Yandex" presented Sona
Sona was trained from scratch based on "Yandex" data. Engineers from "Alisa" and "Yandex Music", as well as the company's research and development team, participated in its development and testing.
In traditional recommendation systems, tasks are divided among several algorithms. Some of them find suitable options from millions of choices, while others evaluate them and determine the order in which they are displayed.
Sona, on the other hand, performs the selection and ranking of options using a single model.
### The model learns from previous choices
Sona analyzes what the user has previously chosen and what content they have returned to. Based on this data, it generates recommendations that may suit the user.
The developers note that this eliminates the need to manually select the features the model relies on.
Using a single model also makes it easier to update the recommendation system. While the traditional approach requires retraining individual algorithms and coordinating them with each other, Sona combines these tasks into a single system.
### What were the results of the one-week test?
Engineers tested Sona for music recommendations on smart speakers with "Alisa". In the test group, the usual system consisting of dozens of algorithms was replaced with the new model for one week.
According to the results presented by the company: the listening time of recommended songs increased by 6.3 percent; requests to replay a song increased by nearly 18 percent; the active listener audience