Kuaishou Recommendation System: Architecture, CTR Modeling, Multi‑Domain Multi‑Task Learning, and Long‑Short Term Behavior Modeling

Kuaishou's recommendation system operates as a large‑scale information‑retrieval engine without explicit queries, training on massive user behavior logs and feedback to predict user interests.

The pipeline follows a funnel‑like structure: a recall pool containing billions of videos is narrowed by a coarse‑ranking stage to thousands of candidates, refined by a fine‑ranking stage to a few hundred, and finally re‑ranked to present dozens of videos to the user.

Key characteristics include an enormous data volume (hundreds of billions of short videos and billions of users), highly diverse user behavior, rich content variety, and complex interaction scenarios across multiple UI flows such as the main discovery page,精选,极速版,关注页,同城页, and live streams.

The core personalization component is the CTR model, which drives click‑through prediction in dual‑stage interactions. Initial attempts to increase model expressiveness involved a global network combined with a personalized subnet (inspired by ResNet) and a user‑bias approach (inspired by LHUC), but gains were limited due to sparse user‑ID features, unstable embeddings, and static user representations.

PPNet (Parameter‑Personalized Net) was introduced as the final solution. It adds a per‑user and per‑video gating network at every DNN layer, enabling dynamic, personalized networks for each sample. PPNet delivered substantial retention improvements, especially for newly active users, and was later migrated from CPU to GPU for higher inference performance.

To support many products, a multi‑domain multi‑task learning framework was designed. It aligns semantics, embeds space via a gate network, aligns feature importance across domains, and incorporates a personalized MMoE that adds task‑specific bias terms, allowing shared learning while respecting domain differences.

For short‑term behavior sequence modeling, an encoder‑only transformer was adopted, replacing self‑attention with target attention and using video watch timestamps instead of position embeddings, which reduced computational cost and significantly boosted performance.

Long‑term behavior modeling faced challenges of sparse signals and ultra‑long sequences. After evaluating Alibaba's SIM approach, Kuaishou built V1.0 (tag‑based retrieval) and V2.0 (embedding‑distance retrieval) solutions, introducing high‑coverage category systems, clustering, and distance‑based matching to achieve robust long‑term interest representation.

The system scaled feature count to the hundred‑billion level and parameter count to the trillion level, mitigating feature sparsity, improving model convergence, and delivering notable market gains.

Future work will focus on cross‑business model fusion, deeper multi‑task learning, unified short‑ and long‑term interest modeling, and advanced user retention prediction.