Kuaishou Recommendation System: Precision Ranking Model Practices and Multi‑Task Learning

01. Kuaishou Recommendation System Kuaishou's recommendation pipeline follows an information‑retrieval paradigm without an explicit user query. Billions of short videos form a candidate pool; the recall stage selects thousands for coarse ranking, then hundreds for fine ranking, and finally dozens for re‑ranking. The platform serves over 300 million users with diverse interests and massive daily video deliveries.

02. CTR Model – PPNet PPNet, introduced in 2019, is a personalized click‑through‑rate estimator built on a fully‑connected DNN. To capture user‑specific patterns, the team experimented with stacking user‑specific subnetworks and later added per‑user bias and vector terms inspired by LHUC. Although gains were modest, a gated network combined with a global baseline yielded noticeable improvements, especially for sparse‑behavior users, and was later accelerated with GPU inference.

03. Multi‑Domain Multi‑Task Learning Framework Kuaishou serves many front‑end scenarios (discovery pages,精选,极速版, etc.) and user groups, resulting in dozens of prediction targets. To avoid inefficient isolated models, a unified multi‑task architecture was built using feature semantic alignment, embedding transform gates, slot gates for feature importance, and a shared MMOE tower with personalized bias. This unified model increased overall interaction metrics by roughly 10%.

04. Short‑Term Behavior Sequence Modeling For short‑term user actions, the team replaced RNNs with sum‑pooling and introduced four enhancements: (1) an encoder to represent historical sequences, (2) inclusion of video‑play duration and interaction labels, (3) replacing self‑attention with target‑attention using current embeddings, and (4) using log‑time‑difference instead of positional embeddings. These changes delivered substantial online gains.

05. Long‑Term Behavior Sequence Modeling To capture users' long‑term interests, two versions were developed. V1 (tag‑based retrieval) stored years of user behavior on high‑density AEP storage, applied back‑track completion and max‑path matching, and optimized Transformer costs. V2 (embedding‑distance retrieval) clustered video embeddings and performed cosine similarity search. Both versions markedly improved user retention and app usage time.

06. Billion‑Feature Trillion‑Parameter Model The team identified feature sparsity and memory pressure as bottlenecks and upgraded the parameter server (GSET) with better memory control, custom feature‑score eviction (outperforming LFU/LRU), and integration of Intel AEP non‑volatile memory via the NVMKV engine, enabling stable training of models with billions of features and trillions of parameters.

07. Summary and Outlook Future work will focus on deeper model fusion, advanced multi‑task learning, and better integration of short‑ and long‑term interest modeling as well as user retention prediction.