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

Overview With the rapid growth of big data and deep learning, health‑e‑commerce recommendation systems are increasingly incorporating large‑model (LLM) technologies into the traditional "recall‑ranking" pipeline, either as augmentation or replacement. The article outlines JD Health's approach to merging large models with recommendation pipelines and examines domain‑specific challenges.

1. Large‑Model Recommendation Technology Review Three main technical directions are identified: (1) scaling traditional CTR models to larger sizes, (2) enhancing existing stages with LLMs for richer sample generation, feature enrichment, and embedding improvements, and (3) adopting generative large models for end‑to‑end recommendation, exemplified by Meta's Trillion‑Parameter Sequential Transducers.

2. Health E‑commerce Recommendation Background and Challenges Health products exhibit strong demand‑driven and knowledge‑driven characteristics, with many items being standardized (e.g., masks, vitamins). Challenges include sparse user behavior, low‑frequency purchases, and the need to balance knowledge‑based recommendations with SKU‑level personalization.

3. Large‑Model Deployment in E‑commerce

3.1 LLM4CB – Solving Sparse‑User Recall

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

Solution: Leverage LLM world knowledge to infer potential needs, combine with domain samples, and optimize inference latency via offline‑to‑near‑line pipelines.

Technical challenges: material representation conversion (ID vs. token), task alignment, and inference efficiency.

Implementation: Two‑stage modeling – first generate potential product names (PU) using prompt‑engineered LLMs, then recall specific SKUs via scoring (EE) mechanisms such as UCB or Thompson Sampling.

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

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

3.2 DeepI2I – Extending Item‑to‑Item Models To address long‑tail items, the DeepI2I model expands sample size, parameters, and epochs, using graph neural network random walks and LLM‑based data augmentation to improve low‑frequency item coverage.

3.3 Large‑Model CTR – Practicing Scaling Law CTR models are scaled in depth, width, and multimodal inputs, following the scaling‑law principle to enhance feature interaction and prediction accuracy for complex health‑e‑commerce scenarios.

4. Open Questions and Personal Viewpoints The article questions whether generative recommendation can fully replace the multi‑stage filtering paradigm, emphasizing that e‑commerce recommendation remains a systems problem involving supply, distribution, and marketing. It argues that combining scaling‑law benefits with LLM knowledge offers the most promising path forward.

Conclusion JD Health's recommendation team demonstrates that large‑model techniques can effectively enrich sparse‑user recall, improve relevance, and drive performance gains in health e‑commerce, while highlighting ongoing research needs in model alignment, infrastructure scaling, and generative recommendation paradigms.