The VK AI team has developed a new method for predicting video watch time called Hierarchical Exponential-Gaussian Mixture (HEGM). This approach addresses the limitations of standard regression models, which often fail to capture the multimodal nature of user behavior in short-video feeds, such as quick skips versus full views.
Standard regression models, which optimize for Mean Squared Error or Mean Absolute Error, often produce a single point estimate that falls between distinct behavioral modes. This results in a loss of information regarding the probability of specific actions like quick skips or complete video completions.
The HEGM model improves upon previous flat mixture architectures by using a two-level hierarchy. A sigmoid gate first estimates the probability of a quick skip, while a separate branch of Gaussian components models meaningful viewing behavior. This structure prevents competition between the skip and view modes.
To ensure stability and prevent model degeneration, the team implemented a specific initialization strategy. Gaussian component centers are distributed across the duration range, and an entropy-based prior on component width is applied to prevent the model from collapsing into narrow, uninformative spikes.
The research has been accepted for the main track of the IEEE ICDM 2026 conference. The VK team has released the code and models for the HEGM approach to the public, providing a more robust alternative to traditional point-based forecasting for recommendation systems.
