Evolution and Architecture of JD.com’s Personalized Recommendation System

In e‑commerce, recommendation creates value by uncovering latent purchase intent, shortening the distance between users and items, and enhancing the shopping experience.

JD.com’s recommendation started in 2012 with simple rule‑based matching and evolved through several phases: a fragmented early stage, a 2013 redesign triggered by the big‑data era, a 2015 upgrade to support multi‑screen (App, PC, mobile web, WeChat, etc.) and diverse recommendation types such as activities, coupons, and articles, and a 2016 breakthrough with the “Intelligent Mall” that dramatically boosted GMV and reduced manual costs.

The recommendation product line now spans multiple screens and types, integrating user behavior tracking via front‑end instrumentation, click‑stream collection, and real‑time feature computation to enable precise personalization across devices.

The overall business architecture provides a unified HTTP recommendation service for all terminals, supported by model services (user and item profiling), machine‑learning model training, and a data platform that handles both offline batch processing (MapReduce/Spark) and online real‑time computation (Kafka, Storm/Spark, Redis, HBase).

The personalized recommendation architecture decouples data, algorithms, and interaction layers, aiming to transform “one‑size‑fits‑all” into “one‑size‑fits‑each” through data mining, machine learning, and scenario‑aware re‑ranking, supporting multiple recommendation types and A/B experimentation.

The data platform collects user actions from all JD touchpoints via click‑stream, stores logs for offline model training, and feeds real‑time streams to online services; offline pipelines generate features, user/item portraits, and models on Hadoop/Spark, while online pipelines compute real‑time behavior, portraits, and feedback using Kafka, Storm/Spark, and store results in Redis/HBase.

The recommendation engine follows a classic pipeline: candidate recall (based on user/item profiles, cold‑start services for new users), rule‑based filtering, feature computation (including real‑time behavior and knowledge‑graph signals), scoring with machine‑learning models, and diversified merging of results from multiple recommenders.

An example is JD App’s “Guess You Like” feature: user portraits and recent behavior select recall strategies, business rules filter candidates, cross‑features are extracted, models score items, and the final list is re‑ranked, enriched with recommendation reasons, and adjusted for diversity.