Real-time Feature Engineering for Recommendation Systems via Edge Computing

In recent years, rapid development of cloud computing and big data has accelerated machine‑learning‑based recommendation systems. Cloud‑based centralized computation can update models daily (or hourly during peak events) but suffers from minute‑level latency when processing massive behavior data.

Real‑time performance is critical: model freshness captures emerging trends, while feature freshness reflects individual user actions. Delays in feature updates cause users to lose interest.

Two aspects of real‑time recommendation are emphasized: the timeliness of the model and the timeliness of the features. Model updates can be done daily, but feature updates must react within seconds to user behavior.

The article lists real‑time features used in Xianyu, such as browsing exposure, exposure duration, scroll speed, detail‑page actions (click, like, comment), and purchase actions.

To overcome cloud latency, edge (client‑side) computation is introduced. Edge devices collect richer behavior dimensions and perform real‑time clustering and intent inference before reporting to the cloud.

Real‑time clustering aggregates events in 60‑second windows, counting exposures, durations, and clicks per category. Strong events (clicks) trigger immediate reporting; weak events (exposures) are batched.

Real‑time intent features use an on‑device intelligent model to interpret user intent from behavior streams and send the results upstream.

By moving feature engineering to the edge, the system achieves decentralization, reduces latency, improves recommendation accuracy, and demonstrates that edge‑cloud collaboration is a promising future direction.