Industrial-Scale Graph Learning for JD Advertising: 9N GRAPH End‑to‑End Solution and BVSHG Model

JD.com's 9N algorithm framework is widely used in recommendation and search advertising. The article explains how large‑scale graph algorithms are deployed in JD's advertising scenarios, starting with an analysis of e‑commerce recommendation challenges such as sparse user behavior, massive item catalogs, fast‑changing interests, and hot‑item updates.

To address these issues, a user‑item bipartite graph is constructed to capture heterogeneous relationships between users, items, categories, and brands. This graph enables the integration of long‑tail and highly active user data, mitigating data distortion caused by truncation or sparsity.

The 9N GRAPH industrial solution consists of six layers, including data & sample generation, online graph services with second‑level updates, a training framework (9N Lite → 9N GL), algorithm modeling, and deployment. Key challenges tackled are massive graph sample storage, real‑time graph service, and joint training of main and graph models.

9N Lite provides a TensorFlow‑based, elastic training engine that supports supervised learning, reinforcement learning, online learning, federated learning, and graph learning. It offers high‑performance data flow, operator libraries, a core training framework, and a model management platform (NormGuard).

The BVSHG model (Multi‑behavior Multi‑view Session‑based Heterogeneous GNN) is introduced as a concrete end‑to‑end solution. It builds a large heterogeneous graph with user, item, category, and brand nodes, encodes multiple behavior types (click, purchase, share, add‑to‑cart, etc.), and samples neighbors using behavior‑weighted random and homogeneous sampling.

BVSHG combines three embeddings: a traditional CTR model (user, item, context features), a graph model embedding, and a short‑term session transformer embedding. Multi‑view attention aggregates item‑item, item‑category, and item‑brand relations, and the final concatenated representation is fed to an MLP for pCTR prediction.

Experimental results on JD's recommendation advertising show increased click‑through and conversion rates, demonstrating the effectiveness of the graph‑enhanced end‑to‑end training. Future work includes AutoML for graph hyper‑parameter tuning and graph‑based reinforcement learning to capture long‑term user value.