Integrating Large Language Models into Health E‑commerce Recommendation Systems: Development, Challenges, and Practice

The rapid growth of big data and deep‑learning has prompted e‑commerce platforms to embed large‑model technologies into the classic "recall‑ranking" pipeline. This piece outlines how JD Health’s recommendation team combines large‑model methods with existing systems, focusing on three core topics: a historical review of large‑model recommendation techniques, the unique background and challenges of health‑e‑commerce recommendation, and concrete deployment practices in the e‑commerce domain.

1. Review of Large‑Model Recommendation Techniques

Three main technical directions are identified:

Traditional CTR models scaling up to massive architectures (Wide & Deep, DeepFM, xDeepFM) and leveraging distributed training and GPU acceleration.

LLM‑enhanced recommendation, where large language models generate high‑quality product descriptions, enrich user profiles, and improve embedding representations for better personalization.

Generative large models that shift the paradigm from multi‑stage filtering to end‑to‑end generation, exemplified by Meta’s Trillion‑Parameter Sequential Transducers for Generative Recommendations.

2. Health E‑commerce Recommendation Background and Challenges

Health products exhibit strong demand and knowledge‑driven purchase behavior, often involving standardized items (e.g., vitamins, masks) where user decisions depend more on medical knowledge than on diverse interests. Challenges include sparse purchase data, user‑behavior sparsity, and large variations across usage scenarios, requiring fine‑grained algorithms and deep user understanding.

3. Practical Deployments in E‑commerce

LLM4CB – Solving Sparse‑Behavior Recall

Problem: New or low‑frequency users lack sufficient interaction data for traditional recall.

Solution: Use LLM world knowledge to infer potential needs from basic user attributes, combine with domain‑specific samples for task alignment, and optimize inference latency via offline‑to‑near‑line pipelines.

Technical challenges: Mapping item IDs to LLM token representations, aligning recommendation tasks with LLM objectives, and ensuring real‑time inference performance.

Implementation: A two‑stage workflow—first generate candidate product names (PU) via prompt‑engineered LLM, then recall concrete SKUs using an EE scoring function.

Mathematical formulation:

\[ \text{SU} = \{ \text{sku} \mid \text{EE}(\text{pu}, \text{sku}) \geq \epsilon \} \]

\[ \text{pu} = G(\text{u, cxt}) \mid f(\text{P, G}) \geq \theta \]

where SU denotes the set of SKUs recalled for user u , EE is the scoring function, pu is the generated product set, and f measures the match between generated and real products.

DeepI2I – Extending Item‑to‑Item Models

To address long‑tail recommendation, the team augments the classic I2I model with increased data volume, larger parameter counts, and more training epochs, employing graph neural network random walks for low‑frequency item sampling and LLM‑based data augmentation.

Large‑Model CTR – Practicing Scaling Law

CTR models are scaled up in complexity (multi‑modal, multi‑objective) and computational intensity, following the scaling‑law principle that larger models with more data improve performance. The system transitions from storage‑centric to compute‑centric architectures to handle richer feature interactions.

Open Questions and Personal Views

Can generative recommendation fully replace the traditional multi‑stage filtering paradigm at scale?

Recommendation systems are interdisciplinary, involving supply chain, distribution, and marketing, not just algorithms.

Large models provide significant gains but should be combined with existing pipelines rather than discarded.

In summary, JD Health’s large‑model initiatives demonstrate how world knowledge and domain‑specific data can jointly improve cold‑start recall, how scaling‑law‑driven model expansion enhances CTR prediction, and how careful system engineering (prompt design, asynchronous inference, KV storage) maintains real‑time performance.