Kuaishou's HGenPush boosts DAU 0.181% with hybrid gen AI
Kuaishou deployed a new gen AI recommender that recommends both videos and authors in one model.
Traditional recommendation systems struggle to serve both content discovery and creator loyalty in short-video platforms. Users want high-quality videos and trusted authors, but existing generative recommenders are limited to single-type outputs and rely on slow autoregressive generation of semantic IDs. These inefficiencies reduce scalability for real-time push notifications.
To solve this, Kuaishou's HGenPush introduces an end-to-end heterogeneous generative architecture. It features a hybrid user behavior understanding module that fuses multi-scenario and multi-perspective actions to capture precise interests. A dual-branch generative module recommends both videos and authors simultaneously. Crucially, a lightweight multi-token prediction method replaces autoregressive token generation, speeding up inference. A consumption preference alignment module uses user feedback as reward signals to guide the model toward higher-quality content. Deployed on Kuaishou's push notification system, HGenPush lifted daily active users by 0.181%, proving its real-world impact.
- Combines video and author recommendation in a single generative model, addressing both content quality and creator trust.
- Uses multi-token prediction instead of autoregressive generation, improving efficiency for real-time push notifications.
- Deployed at Kuaishou's push system, achieving a 0.181% increase in DAU, a significant lift for a platform of its scale.
Why It Matters
This architecture shows how gen AI can make push notifications more relevant by recommending both content and creators.