Long-Term User Interest Modeling for Click‑Through Rate Prediction in Alibaba’s Advertising System

CTR estimation is a core technology for advertising and recommendation, and modeling long‑term user interests requires processing extensive historical behavior sequences, which creates storage and latency challenges for online services.

The authors first outline the business context of Alibaba‑Mama’s targeted advertising and review the evolution of their models from LR to MLR, DNN, DIN, and DIEN, highlighting that longer behavior windows (e.g., 120 days) dramatically increase sequence length and expose system bottlenecks.

To overcome these bottlenecks, a co‑design of algorithm and system is proposed with three breakthroughs: computation decoupling, storage compression, and incremental reuse. A Memory Network is adopted as the base model because it can capture long‑term dependencies, reduce storage pressure, and lower computational cost compared with GRU‑based approaches.

Two novel memory‑network variants are introduced: Multi‑channel User Interest Memory Network (MIMN), which aggregates multi‑peak interests via content‑based addressing and a Memory Induction Unit, and Hierarchical Periodic Memory Network (HPMN), which models interests across short, medium, and long time scales using a hierarchical RNN‑like structure with diversity regularization.

System-wise, an asynchronous computation module updates interest vectors independently, and a User Interest Center (UIC) stores compressed interest vectors for fast online retrieval. The redesigned pipeline replaces raw behavior features with interest vectors, enabling real‑time attention and incremental updates.

Extensive offline experiments on Amazon and Taobao datasets show that both MIMN and HPMN outperform baselines (Embedding & MLP, DIN, DIEN, etc.) in GAUC, with up to 0.6% improvement when extending behavior length from 100 to 1000. Online A/B tests report 7.5% CTR lift in “Guess You Like” and 6% RPM increase.

Practical lessons include handling model parameter synchronization between UIC and RTP servers, mitigating promotion‑period data bias, initializing long‑term interest states, and implementing rollback mechanisms for fault tolerance.

Overall, the work demonstrates that jointly designing algorithms and system infrastructure can unlock the value of long‑term user behavior for more accurate CTR prediction.