Yandex Implements a Unified Neural Network for Music Recommendations

In services with personalized feeds, recommendations are typically generated using complex chains of multiple models. Yandex conducted an experiment to test the feasibility of replacing this entire system with a single neural network.

Previously, dozens of algorithms were used to select music in smart speakers with Alice. Engineers manually tuned features for track evaluation, and each stage of the cascade was trained separately.

The new Sona model works like a language model but predicts the next track instead of a word. A code consisting of three tokens is created for each song based on the first 90 seconds of audio and metadata.

• The model receives raw user action history of up to the last 8192 records
• It independently discovers patterns without manual feature engineering
• It receives data on skipped tracks for retraining approximately every 45 minutes

A week-long test showed a 6.3% increase in listening time for recommended tracks. The number of requests to repeat a track increased by nearly 18% compared to the old system.

Major tech companies are actively adopting unified recommendation models. Chinese Kuaishou replaced its chain with a single model, Pinterest merged some stages, but Yandex became the first to test a full replacement with autonomous criterion selection.

The success of the experiment confirms the effectiveness of this approach. Yandex plans to develop recommendation systems based on unified neural networks to improve user service quality.