The VK Research team has introduced HEGM, a machine learning method designed to improve content recommendation accuracy. The technology is already integrated into the VK Clips platform to better predict user engagement.
Following the implementation, the number of videos watched per session rose by 10 percent. Additionally, the volume of views lasting longer than 10 seconds increased by 6 percent, while negative feedback dropped by 8 percent.
Unlike standard algorithms that rely on average interaction time, HEGM analyzes various user behaviors, including skipping, full viewing, and replaying. It calculates the probability distribution of interaction duration to forecast multiple engagement metrics simultaneously.
The model was validated using the KuaiRec and VK-LSVD datasets, as well as an internal sample of 282 million anonymized user interactions. HEGM demonstrated superior ranking performance compared to existing solutions.
The research has been accepted for presentation at the IEEE ICDM 2026 conference. Developers can access the source code for the model on GitHub for potential applications in music streaming, advertising, and e-commerce.
