How JD Leverages Knowledge Graphs for Better E‑commerce Interest Recall

Introduction

Peng Changping from JD Recommendation shares algorithmic innovations developed over the past one to two years in JD’s e‑commerce recommendation system.

Why E‑commerce Recommendation Differs

E‑commerce recommendation must both capture user interests and expand them; interest expansion accounts for more than half of the exposure in JD’s system.

Challenges of Interest Expansion

Key challenges include sparse user behavior, billions of SKU candidates, cold‑start new products, promotional spikes, time‑sensitive hotspots, and bundle‑purchase scenarios, all of which make pure behavior‑based methods ineffective.

Knowledge‑Graph‑Based Interest Recall

JD builds a product knowledge graph with four layers: raw SKU items, extracted entities (from titles, descriptions, reviews), a concept layer (key product attributes), and the final product level. This graph enables recall of items even when behavior data are missing.

The recall‑ranking funnel consists of three stages:

Recall thousands of candidate SKUs from billions using the knowledge graph.

Estimate click‑through‑rate (CTR) for each candidate.

Rank the final list, mixing items that match known interests with those discovered through expansion.

CTR Estimation with DRM Module

To improve CTR prediction for long‑tail and cold‑start items, JD introduces a DRM (Dimension‑Relation‑Model) module. It applies self‑attention over embedding matrices, refining sparse features and preventing over‑fitting, which yields higher AUC and lower loss across embedding dimensions.

Listwise Ranking and Exploration

After CTR and CVR estimation, JD generates multiple candidate sequences using heuristic and clustering‑based exploration, then applies a listwise scoring model that jointly optimizes click, GMV, and diversity. This approach maintains click performance while significantly improving result diversity and user dwell time.

Conclusion

The combination of knowledge‑graph‑driven recall, the DRM‑enhanced CTR model, and listwise ranking with exploration has delivered measurable gains in diversity, click‑through, and revenue for JD’s recommendation platform.