VK Unifies Service Recommendations Using a Single Neuroprofile

VK's artificial intelligence development team has introduced a unified user neuroprofile—a technology that aggregates human behavior across different ecosystem services to generate cross-service personalized recommendations.

How the Architecture Works

The system is based on a two-tower transformer architecture:

• User Tower: processes up to 1,024 of the user's latest actions into a single sequence using a self-attention mechanism, generating a dynamic vector of their current interests.
• Item Tower: encodes the characteristics of videos, posts, and clips using multimodal semantic embeddings.

Each event within the history is transformed into an informative token. It accounts for the content vector, action type (like, repost, full view, or skip), post author, as well as temporal factors via the Time2Vec approach (absolute time, current session dynamics, and switching speed between items).

Two-Stage Training and Causal Sessions

To avoid bias toward overly popular content and noise, the training was split into two stages:

1. Pre-training (Pre-train): baseline filtering of irrelevant content based on implicit engagement signals (watch depth) using the Sampled Softmax technique.
2. Fine-tuning (Causal Fine-tune): tuning for target actions (likes, saves, complete views). The user's timeline is divided into distinct sessions by content type with prefix-history isolation. This prevents data leakage from the future and teaches the model to accurately predict intent immediately before a session begins.

The "Heimdall" Cross-Domain Model and Results

To bridge the gap between short and long formats, the "Heimdall" model was developed. The algorithm balances data across services with varying interaction frequencies and predicts the next relevant item rather than isolated micro-actions.

The deployment of the cross-domain neuroprofile demonstrated measurable metric improvements:

• Watch time in VK Clips increased by 5.5%.
• Total View Time (TVT) on the VK Video main page grew by 1.1% on top of existing local models due to cross-service ecosystem context.