How MIMN+UIC Breaks the Long-Sequence Barrier in Real-Time CTR Prediction

Background

Alibaba's precise targeting advertising team observed that users exhibit diverse interests, but only a subset influences behavior for a specific item, which led to the DIN network. To capture abstract user interests and their evolution, they later introduced the DIEN model. However, online systems typically limit sequence length to about 50, missing richer behavior information.

Proposed Solution

From an algorithm‑and‑system co‑design perspective, the authors propose a new CTR prediction model called MIMN (Multi‑channel user Interest Memory Network) and a dedicated UIC (User Interest Computing) module that decouples interest calculation from ad request processing.

Algorithm Details

MIMN reads ultra‑long user behavior sequences, extracts and aggregates diverse interests into a memory network. It incorporates memory utilization regularization to improve storage efficiency and an Interest Induction Unit (MIU) that captures the evolution of multiple interest tracks.

Memory utilization regularization reduces the variance of write weights across memory slots, encouraging balanced memory usage. The model also introduces a memory induction unit that treats each slot as an interest track and updates them with GRU‑based evolution.

System Design

The independent UIC server incrementally updates user interest whenever new behavior arrives and stores the representation in TAIR. The real‑time CTR prediction service retrieves the latest interest vector from TAIR, eliminating additional latency and decoupling interest inference from the prediction request.

Experiments

Extensive experiments were conducted on Amazon (books), Taobao public datasets, and Alibaba's production ad data. Using 1000‑length sequences yields a 0.6% offline AUC improvement over 100‑length sequences. In production, the MIMN+UIC architecture improves online CTR by 7.5% and RPM by 6% compared with the best existing model DIEN.

Deployment Insights

Parameter synchronization between UIC and RTP servers is handled via hourly incremental training, reducing inconsistency risk.

During major promotion periods (e.g., Double 11), data distribution shifts cause a slight offline performance drop (~0.2%).

An initialization strategy exports 120‑day interest vectors to TAIR to accelerate cold‑start.

A rollback mechanism saves daily snapshots of user interest states, enabling recovery from data contamination or system failures.