Generative Recommendation Systems for JD Alliance Advertising: Architecture, Implementation, and Experimental Evaluation

Large language models (LLMs) are profoundly influencing natural language processing and open new research avenues for recommendation systems (RS), especially in e‑commerce advertising such as JD Alliance. This paper first introduces the motivation for combining generative RS with JD Alliance ads, then reviews existing pipelines and presents a practical implementation.

Generative Recommendation System – Unlike traditional multi‑stage RS that rely on recall, coarse ranking, fine ranking and re‑ranking, a generative RS can directly generate recommended items using LLMs, simplifying the pipeline and offering better generalization and stability, particularly for cold‑start and sparse‑data scenarios.

JD Alliance Advertising – JD Alliance is a CPS advertising platform that drives traffic through external links. Its ad recommendation faces challenges such as data sparsity, cold‑start, scene understanding, and maintaining diversity and novelty for low‑activity users.

Integration of LLMs – By leveraging LLMs’ world knowledge and contextual understanding, the system can improve item representation, user modeling, and context incorporation, thereby enhancing recommendation accuracy and relevance.

Four Core Stages – The generative RS workflow consists of (1) material representation, (2) model input formulation, (3) model training, and (4) model inference. Material representation explores numeric IDs, textual metadata, and semantic IDs (SID) obtained via RQ‑VAE quantization. Model input combines task description, user history, user profile, and contextual information. Training optimizes next‑item prediction and alignment tasks between SID and textual descriptions. Inference uses beam search with constrained decoding (e.g., Trie) to ensure generated IDs map to real items.

Material Representation Details – Numeric IDs are tokenized into sequences; textual metadata provides semantic cues but suffers from length and ambiguity; semantic IDs are derived from vector quantization of BERT/Yi‑6B embeddings, optionally disambiguated with extra tokens. The paper includes a comparison table of these methods.

Model Input – Task prompts guide the LLM to perform next‑item prediction, while user interaction sequences are encoded as tokenized item IDs. User profiles and contextual signals (e.g., time, location) are appended to enrich the prompt.

Training and Inference – Base models such as Qwen1.5 (0.5B/1.8B/4B) and Yi‑6B are fine‑tuned with LoRA and multi‑task objectives. Inference employs beam search (beam size = 20) with optional constrained decoding to guarantee valid item IDs.

Experimental Results – Offline experiments show larger models achieve higher HR@K and NDCG@K, with Yi‑6B outperforming collaborative baselines. Online A/B tests on low‑activity users demonstrate that the generative RS matches or exceeds traditional multi‑recall pipelines, achieving >5% UCTR improvement on certain pages.

Optimization Directions – Future work includes scalable SID training frameworks, multi‑LoRA and mixed‑data strategies, model distillation, pruning, quantization, and extending the system to handle search‑query integration, recommendation explanations, and preference generation.

References – The article cites over 30 recent works on LLM‑based recommendation, semantic ID quantization, and e‑commerce applications.

{
    "instruction": "该用户为都市女性。用户已按时间顺序点击了如下商品：
,
, ...，你能预测用户下一个可能点击的商品吗？",
    "response": "
"
}

{
    "instruction": "商品
的标题是什么？",
    "response": "ThinkPad 联想ThinkBook 14+ 2024 14.5英寸轻薄本..."
}

{
    "instruction": "哪个商品的标题是\"ss109威震天变形MP威震玩具天金刚飞机威男孩机器人战机模型合金 震天战机（战损涂装版）\"？",
    "response": "
"
}