Large-Scale Graph Technology in JD.com E‑commerce: Practice and AI Computing Directions

On October 26 at the AICC 2021 Artificial Intelligence Computing Conference, JD.com Vice President Bao Yongjun delivered a keynote titled “Large‑Scale Graph Technology in JD.com E‑commerce Practice” and later answered questions about the future direction of AI computing and its concrete applications within JD.

As JD’s retail business expands, artificial intelligence has become a key driver for growth, underpinning marketing, fine‑grained operations, and user interaction, all of which increasingly rely on graph technology for precise data modeling and complex relationship representation.

In JD’s e‑commerce scenario, product and user graphs form a solid data foundation that supports two major business categories: (1) user‑centric operations such as search, recommendation, and advertising, and (2) risk‑control functions like fraud detection. The product graph stores massive items and their static and dynamic relationships, while the user graph captures multi‑dimensional user profiles derived from interactions with products.

Risk‑control leverages graph representations of malicious groups, enabling the detection of coordinated fraud behaviors; JD reports that graph‑based detection improves effectiveness by 2.55 times. However, deploying graph computing at JD’s scale faces challenges in storage efficiency, millisecond‑level response latency, and the need to support advanced algorithms on heterogeneous data.

To meet these demands, JD adopts the Galileo graph engine, building heterogeneous multi‑type graphs that integrate user attributes, interaction behaviors (view, add‑to‑cart, share, like, etc.), and product taxonomy information. This enables both short‑term sequential graph modeling of user actions and long‑term behavior modeling, facilitating richer semantic representations and more accurate interest inference.

In the interview, Bao outlined four layers of AI computing development: (1) chip technology to boost AI compute power, especially amid US‑China trade tensions; (2) automated learning systems that enable end‑to‑end AI model generation; (3) application‑level AI across text understanding, computer vision, automated risk control, and emerging media such as video and speech; and (4) data privacy and compliance, ensuring secure, lawful use of user data across JD’s ecosystem.