Hierarchical Exponential-Gaussian Mixtures for Watch-Time Distribution Prediction
The Hierarchical Exponential-Gaussian Mixture (HEGM) model improves existing Exponential-Gaussian Mixture Network models for predicting watch time in short videos. It addresses failures like variance collapse and component redundancy by introducing a hierarchical skip-watch decomposition, KL-based variance regularization, and structured initialization. This results in a more stable, interpretable, and accurate distributional model for user behavior.