VK Team Developed AI Model to Improve Short Video Recommendations
The new HEGM architecture enables more accurate prediction of video watch time and accounts for different user behavior scenarios.
Researchers from VK have introduced HEGM (Hierarchical Exponential-Gaussian Mixture), a machine learning model for predicting the watch time of short videos. The development aims to enhance the accuracy of recommendation systems by analyzing various scenarios of user interaction with content.
Unlike traditional methods that predict watch time with a single number, HEGM models an entire probability distribution. This allows it to distinguish between quick skims, partial views, full completions, and repeat plays.
The new architecture also addresses training issues in previous probabilistic models, including variance collapse and the deactivation of individual components. Based on the results of experiments and A/B testing on real traffic, the researchers evaluated the effectiveness of the proposed approach.
The work has been accepted into the main track of the IEEE ICDM 2026 conference. The team has published the code and trained models in open access.
