How JD’s Recommendation System Fuels Business Growth: Insights from Peng Changping

How JD’s Recommendation System Drives Business Growth

Interview guest Peng Changping, graduate of the Chinese Academy of Sciences, with over ten years of experience in machine learning and recommendation systems, now leads JD’s recommendation advertising algorithms.

Recommendation systems are ubiquitous in e‑commerce and serve as major revenue engines. JD’s personalized recommendation system has generated significant profit, and the team explores large‑scale machine learning, deep learning, and other techniques to improve search and recommendation.

What Makes a High‑Quality Recommendation System?

Personalized recommendation meets the massive “thousands of users, thousands of items” demand by building user interest models using machine‑learning or deep‑learning algorithms that combine user, item, and context features, thereby shortening the distance between users and products.

Quantity: The SKU count far exceeds human processing capacity; for example, JD carries hundreds of thousands of jam SKUs, and a “Less Is More” approach helps filter suitable items for each user.

Quality: JD’s system promotes items that are “good, cheap, fast,” integrating brand, attribute, price, rating, and logistics information, which improves user experience and platform loyalty.

Large‑scale machine learning and deep learning have become the most widely applied and successful AI technologies in industry, replacing manual decision‑making in many stages of recommendation.

Key Application Areas

Click‑through rate and conversion prediction.

Recall using multiple models: vector‑based, tree‑based, graph‑based.

Item knowledge graphs leveraging NLP and computer‑vision for text, image, and video understanding.

Rerank with multi‑objective optimization and reinforcement learning for session‑level improvement.

Characteristics of an Effective System

An effective system must satisfy user needs (long dwell time), be growth‑oriented (expand interests, support new users and merchants), and reflect platform values (promote healthy competition).

To achieve this, the system must learn from user feedback and item information, employ diverse recall algorithms, and optimize multi‑objective loss functions.

Practical Evolution at JD

JD’s recommendation system has progressed through four stages:

Meeting basic user demand with item‑based collaborative filtering.

Expanding user demand via diverse recall (“recall kaleidoscope”) and richer ranking objectives.

Session‑level global optimization and ecosystem optimization, treating a user’s session as a list for reranking.

Cross‑user‑group and cross‑merchant joint optimization, leveraging knowledge graphs and transfer learning to handle diverse user groups and sub‑scenes.

Key breakthroughs include diversified recall, listwise ranking, and merchant ecosystem optimization, which together improve precision, coverage, and novelty.

Technical Challenges and Future Directions

The biggest obstacle is data: acquiring comprehensive, real‑time data and enabling models to learn efficiently from massive datasets. Solutions involve full‑channel data fusion, near‑real‑time streaming, and combining edge and cloud computing.

AutoML and neural architecture search are emerging to automate model design, reducing reliance on manual tuning.

Looking ahead, JD will focus on content‑driven recommendation (live‑stream shopping), scene‑based recommendation (virtual store layouts), and stronger ecosystem optimization to promote high‑quality merchants and items.