JD Advertising R&D: AI‑Driven Solutions for Traffic Valuation, Multimodal Understanding, Auction Mechanisms, Generative Recommendation, and Large‑Model Engineering

JD Advertising R&D focuses on applying cutting‑edge AI algorithms to improve traffic valuation, multimodal content understanding, auction mechanisms, generative recommendation, and large‑model infrastructure.

Traffic value estimation – describes query intent recognition, challenges of multi‑intent queries, long‑tail categories, and a generative‑discriminative model that augments training data for under‑represented categories.

Multimodal representation – outlines a dual‑stream pipeline that extracts text embeddings (BGE‑large‑zh1.5) and image embeddings (ViT‑CLIP‑base), aligns them via CLIP‑based projection, and aggregates them in a recommendation space using a Gate‑GNN.

Auction mechanism (ListVCG) – introduces a reinforcement‑learning‑based sequence auction that approximates the combinatorial 700‑choose‑4 problem, integrates multi‑objective optimization, and uses reward shaping and curriculum RL to improve long‑term revenue.

Multi‑agent RL for bidding – explains joint optimization of bidding and mechanism agents, offline simulation stages, and challenges of modeling platform revenue and advertiser behavior.

Generative recommendation – presents a quantized product representation (RQ‑VAE) and a series of pre‑training, fine‑tuning, and DPO steps that predict the next product a user will browse. Example prompt and code snippets are shown below:

提示词: 请告诉我,商品的四元组表示为{input_turple}的标题是什么？
输入: <a_1><b_2><c_3><d_4>
输出: 华为（HUAWEI）旗舰手机mate60 pro+ 16G+512GB 宣白

Creative generation – describes the RFNet multimodal reliability feedback network for automatic quality assessment of AI‑generated ad images, and its training on a million‑scale labeled dataset (RF1M), published at ECCV 2024.

Large‑model engineering – discusses inference latency constraints (≈100 ms), cost reduction for 1.5 B‑parameter models, hardware‑software co‑design, distributed training, and deployment pipelines to support million‑QPS advertising services.