Multi-Granularity Distribution Modeling for Video Watch Time Prediction via Exponential-Gaussian MN скачать в хорошем качестве

Multi-Granularity Distribution Modeling for Video Watch Time Prediction via Exponential-Gaussian MN

The speaker argues that watch time is a dense, always-available signal for ranking and growth, correlating with DAU. They propose EGMM, which combines an exponential component for near-zero quick skips with multiple Gaussian components for longer, multimodal watch segments. The model outputs mixture weights and component parameters conditioned on user–video context. Training uses maximum likelihood, an entropy term to balance components, and L1 regularization to improve pointwise estimates. EGMM outperforms strong baselines offline and converges faster. Online A/B tests show increases in watch time and reduced divergence from observed distributions, with positive impacts on long-term retention and DAU. The paradigm generalizes to recommender, search, e-commerce, and ads.