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

Click‑through rate (CTR) estimation is a core technology for advertising and recommendation, and modeling long‑term user interests becomes essential when user behavior histories are rich and span long periods.

The authors first outline the business context of Alibaba Mama’s targeted advertising, explaining that ads are served through a pipeline that scores thousands of candidates per request and ranks them by eCPM.

They then discuss the technical status: DIN and DIEN models capture user interests from historical sequences but are limited to about 100 actions due to system constraints, which corresponds to only 14 days of behavior. Extending the sequence to 1,000 actions yields a 0.6% GAUC lift, demonstrating the value of long‑term data.

System bottlenecks are identified: the online RTP engine cannot handle the computational load of longer sequences, and storing extended user histories would explode storage and I/O costs for billions of users.

To overcome these limits, the team adopts a high‑level co‑design approach with three breakthroughs: computation decoupling, storage compression, and incremental reuse.

The core algorithmic innovation is the adoption of Memory Networks. Two variants are presented:

MIMN (Multi‑channel User Interest Memory Network) : aggregates user interests into a small set of memory slots using content‑based addressing, supports unlimited historical actions, and updates memory vectors asynchronously.

HPMN (Hierarchical Periodic Memory Network) : captures interests at multiple time scales (short, medium, long) with a hierarchical RNN‑like structure and a diversity regularization to keep slots distinct.

Both models introduce a Memory Utilization regularization to balance slot usage and a Memory Induction Unit to evolve interest representations.

System design changes include an asynchronous computation module for independent interest vector updates, compressed storage of interest vectors in TAIR, and a User Interest Center (UIC) that serves these vectors to the online RTP engine, enabling real‑time incremental updates.

Extensive offline experiments on Amazon and Taobao datasets show that MIMN and HPMN consistently outperform baselines such as DIN, DIEN, GRU4REC, and others, with improvements of up to 1% GAUC. Online A/B tests report a 7.5% CTR lift in the “Guess You Like” scenario and a 6% RPM increase.

The paper also shares practical deployment lessons: handling model parameter synchronization between UIC and RTP servers, mitigating the impact of promotion‑period data, initializing long‑term interest vectors, and implementing rollback mechanisms.

In summary, the co‑design of algorithms and system architecture enables effective long‑term interest modeling, delivering measurable business value and highlighting the importance of aligning model innovation with scalable infrastructure.