How Large Language Models Are Transforming Health E‑Commerce Recommendations

With the rise of big data and deep learning, e‑commerce recommendation systems are integrating large‑model technologies into the traditional “recall‑ranking” multi‑stage filtering paradigm, either by fusion or replacement. This article describes how JD Health’s recommendation algorithm team combines large‑model techniques with e‑commerce recommendation, and explores specific applications and technical challenges in the health‑e‑commerce domain.

1. Perspective 1: Degree of Change to Existing Recommendations

Currently, there are three main technical approaches for integrating “large models” into recommendation systems: scaling traditional CTR models, enhancing existing recommendation stages with large language models (LLMs), and using generative large models for end‑to‑end recommendation.

Traditional CTR Large Models

CTR models have grown in scale, evolving from simple logistic regression to deep architectures such as Wide & Deep, DeepFM, and xDeepFM, improving feature interaction and non‑linear representation. With larger datasets and distributed/GPU training, model capacity follows the scaling law, driving continuous growth since the industry shift from LR to DNN around 2016.

LLM‑Enhanced Recommendation

Large language models bring extensive world knowledge and strong comprehension, enhancing recommendation systems in data/sample generation, feature engineering, and embedding understanding. LLMs can generate high‑quality product descriptions and user profiles to alleviate cold‑start, enrich textual features, and capture deeper user interests through pre‑training and fine‑tuning.

Generative Large Models

Generative models represent a disruptive shift from multi‑stage filtering to end‑to‑end generation, offering new data processing and content creation capabilities. An example is Meta’s recent “Trillion‑Parameter Sequential Transducers for Generative Recommendations”.

2. Perspective 2: How LLMs Address Historical Recommendation Bottlenecks

Recommendation systems have faced different bottlenecks at various stages.

Stage 1 – Data Volume

Around 2010, data scarcity limited recommendation performance; sparse user IDs, tags, and side information hindered modeling. The explosion of mobile internet in 2013 alleviated data issues, shifting focus to data collection, cleaning, and labeling, combined with tree‑based models like GBDT and XGBoost.