VK Implements Cross-Domain Neuroprofiles for Advanced Content Discovery
VK has announced the development and deployment of an innovative AI model designed to deeply analyze user actions across its diverse services. This model creates a neuroprofile — an anonymized representation of a user’s interests, aggregated from various types of content interactions.
Unified Model for Comprehensive User Understanding
VK’s initiative addresses a critical challenge for recommendation systems within large ecosystems: connecting disparate signals of user behavior. Mikhail Trapeznikov, head of the recommendation technologies group at AI VK, highlighted that users engage with content in multiple ways — watching long videos on VK Video, short clips on VK Clips, or browsing the VKontakte feed. Each of these signals provides only a partial understanding of their interests.
The new neuroprofile model features a unified transformer architecture that collects and processes signals from all VK services. This enables the formation of a holistic and unified representation of user interests, transcending individual product boundaries.
Application and Architecture of the Neuroprofile
This solution is already actively operational in several key company services, including:
- VK Video
- VK Clips
- VKontakte
The cross-domain approach, fundamental to the neuroprofile, allows recommendation algorithms to significantly enhance their effectiveness. The model is trained on billions of content interactions, ensuring high accuracy and relevance in recommendations. This contributes to a more personalized user experience across the entire VK ecosystem.
I’ve been noticing a definite improvement in VK’s recommendations lately, especially between VK Video and my main feed. It feels less disjointed. However, I still find VK Clips recommendations can sometimes go off the rails, showing me things completely unrelated to my usual interests, almost like a separate algorithm. My tip for others is to actively use the ‘not interested’ button on those rogue Clips; it seems to help the system recalibrate faster across the other platforms.