Tech Insight: Selected JD Retail Technology Papers in Artificial Intelligence (2024)

Tech Insight is a JD Retail Technology column that regularly interprets the latest technical papers and research results from JD’s retail technology team. In 2024 the team published ten papers in top AI conferences such as CVPR, SIGIR, WWW, AAAI and IJCAI.

Paper 01 – CVPR 2024

English title: Exploring Region‑Word Alignment in Built‑in Detector for Open‑Vocabulary Object Detection

Chinese title: 探索内置检测器中的区域‑词对齐以实现开放词汇目标检测

Download: https://ieeexplore.ieee.org/document/10656464

Authors: Heng Zhang, Qiuyu Zhao, Linyu Zheng, Hao Zeng, Zhiwei Ge, Tianhao Li, Sulong Xu

Abstract: Open‑vocabulary object detection aims to detect novel categories that were not present during training. Most modern methods learn a visual‑language space from large multimodal corpora and then transfer the knowledge to an off‑the‑shelf detector (e.g., Faster‑RCNN). However, domain gaps cause information loss during transfer, limiting generalisation to new categories. This work proposes BIND, a built‑in detector framework that eliminates the need for module replacement or knowledge transfer. A two‑stage encoder‑decoder training pipeline first learns fine‑grained region‑word alignment, then trains a DETR‑style decoder on labelled detection data. An anchor‑proposal network generates high‑quality proposals, greatly improving detection efficiency.

Paper 02 – SIGIR 2024

English title: A Unified Search and Recommendation Framework based on Multi‑Scenario Learning for Ranking in E‑commerce

Chinese title: 基于多场景学习的搜推联合建模统一框架

Download: https://arxiv.org/abs/2405.10835

Authors: Jinhan Liu, Qiyu Chen, Junjie Xu, Junjie Li, Baoli Li, Sulong Xu

Abstract: Search and recommendation are the two most important scenarios in e‑commerce. Existing multi‑scenario models share parameters to capture task similarity and use task‑specific parameters for differences, which fails to capture fine‑grained distinctions and under‑utilises the full label space. The proposed framework introduces a user‑interest extraction layer and a feature‑generation layer, followed by a global‑label multi‑task layer that jointly models the main and auxiliary tasks. Experiments on industrial datasets show consistent performance gains.

Paper 03 – SIGIR 2024

English title: Optimizing E‑commerce Search: Toward a Generalizable and Rank‑Consistent Pre‑Ranking Model

Chinese title: 优化电商搜索：构建有泛化性和排序一致性的粗排模型

Download: https://arxiv.org/abs/2405.05606

Authors: Enqiang Xu, Yiming Qiu, Junyang Bai, Ping Zhang, Dadong Miao, Songlin Wang, Guoyu Tang, Lin Liu, Mingming Li

Abstract: In large‑scale e‑commerce platforms, the pre‑ranking stage filters billions of items for downstream ranking. Existing methods focus on improving consistency between pre‑ranking and final ranking while neglecting generalisation to long‑tail items. This paper introduces a dual‑objective approach: (1) multiple binary classification tasks enforce rank consistency, and (2) contrastive learning on item representations enhances long‑tail generalisation. Offline AUC and online A/B tests demonstrate significant gains.

Paper 04 – SIGIR 2024

English title: A Preference‑oriented Diversity Model Based on Mutual‑information in Re‑ranking for E‑commerce Search

Chinese title: 京东搜索重排：基于互信息的用户偏好导向模型

Download: https://dl.acm.org/doi/10.1145/3626772.3661359

Authors: Huimu Wang, Mingming Li, Dadong Miao, Songlin Wang, Guoyu Tang, Lin Liu, Sulong Xu, Jinghe Hu

Abstract: Re‑ranking aims to reorder items by considering inter‑item relationships. Existing methods either sacrifice diversity for accuracy or vice‑versa. This work proposes PODM‑MI, a mutual‑information‑based model that simultaneously optimises relevance and diversity using a variational Gaussian representation of user preferences and a utility matrix derived from mutual information. Experiments on JD’s main search platform show substantial improvements in both offline metrics and online A/B tests.

Paper 05 – ACM (JDivPS)

English title: JDivPS: A Diversified Product Search Dataset

Chinese title: 基于京东电商平台的多样化产品搜索数据集

Download: https://dl.acm.org/doi/10.1145/3626772.3657888

Authors: Zhirui Deng, Zhicheng Dou, Yutao Zhu, Xubo Qin, Pengchao Cheng, Jiangxu Wu, Hao Wang

Abstract: Existing diversified product‑search datasets suffer from limited accessibility and lack of manually annotated user intent. JDivPS is the first publicly available dataset with human‑annotated intents, containing 10,000 queries and ~1.68 M unique products from JD.com, each query annotated with an average of ten intents. Extensive evaluation of diversified ranking models on JDivPS provides a solid benchmark for future research.

Paper 06 – WWW 2024

English title: A Semi‑supervised Multi‑channel Graph Convolutional Network for Query Classification in E‑commerce

Chinese title: 基于半监督多通道图神经网络的类目预估方法

Download: https://arxiv.org/abs/2408.01928

Authors: Chunyuan Yuan, Ming Pang, Zheng Fang, Xue Jiang, Changping Peng, Zhangang Lin

Abstract: Query intent classification is crucial for e‑commerce search. Existing methods rely on click‑based supervision, leading to severe class imbalance and unstable labels for long‑tail categories. The proposed SMGCN leverages similarity scores to expand category information, constructs co‑occurrence and semantic graphs, and employs variational inference to maximise mutual information between user preferences and candidate items. Offline and online experiments show SMGCN outperforms the industry‑best baseline.

Paper 07 – AAAI 2024

English title: PPM: A Pre‑trained Plug‑in Model for Click‑through Rate Prediction

Chinese title: 基于预训练的插件式CTR预估模型

Download: https://arxiv.org/abs/2403.10049

Authors: Yuanbo Gao, Peng Lin, Dongyue Wang, Feng Mei, Xiwei Zhao, Sulong Xu, Jinghe Hu

Abstract: Traditional ID‑based CTR models (IDRec) suffer from poor generalisation, especially for long‑tail items. This work separates ID and modality fusion to the model layer, introducing MoRec (a pre‑trained multimodal encoder trained with CTR supervision) and a Unified Ranking Model (URM) that jointly trains MoRec and IDRec end‑to‑end. Experiments demonstrate improved performance on both offline AUC and online conversion metrics.

Paper 08 – AAAI 2024

English title: Parallel Ranking of Ads and Creatives in Real‑Time Advertising Systems

Chinese title: 京东创意优选：广告商品排序和广告创意优选的并行排序实践

Download: https://arxiv.org/abs/2312.12750

Authors: Zhiguang Yang, Lu Wang, Chun Gan, Liufang Sang, Haoran Wang, Wenlong Chen, Jie He, Changping Peng, Zhangang Lin, Jingping Shao

Abstract: Existing ad‑ranking pipelines place product ranking and creative selection in a serial fashion, limiting the complexity of creative models due to latency constraints. This paper proposes a parallel architecture where creative selection and product ranking run concurrently, sharing inference time budget. Offline joint training and online A/B tests on JD’s ad platform show significant gains in both relevance and creative diversity without added latency.

Paper 09 – AAAI 2024

English title: Generalize for Future: Slow and Fast Trajectory Learning for CTR Prediction

Chinese title: 面向未来的泛化：用于点击率预测的慢速和快速轨迹学习

Download: https://ojs.aaai.org/index.php/AAAI/article/view/27797

Authors: Jian Zhu, Congcong Liu, Xue Jiang, Changping Peng, Zhangang Lin, Jingping Shao

Abstract: Deep CTR models assume i.i.d. data, which is violated in online learning due to domain drift. The proposed “slow‑fast trajectory learning” framework introduces three complementary learners (working, fast, slow) and a trajectory loss to mitigate drift and improve temporal adaptation. Experiments on JD’s live CTR system demonstrate reduced drift and higher prediction stability.

Paper 10 – IJCAI 2024

English title: TaD: A Plug‑and‑Play Task‑Aware Decoding Method to Better Adapt LLMs on Downstream Tasks

Chinese title: TaD+RAG‑缓解大模型“幻觉”的组合新疗法

Download: https://www.ijcai.org/proceedings/2024/728

Authors: Xinhao Xu, Hui Chen, Zijia Lin, Jungong Han, Lixing Gong, Guoxin Wang, Yongjun Bao, Guiguang Ding

Abstract: Large language models (LLMs) exhibit hallucinations despite impressive capabilities. This work introduces Task‑aware Decoding (TaD), a plug‑and‑play method that compares supervised‑fine‑tuned outputs with pre‑fine‑tuned outputs to suppress hallucinations. Combined with Retrieval‑Augmented Generation (RAG), TaD+RAG achieves state‑of‑the‑art performance on knowledge‑intensive QA tasks, demonstrating strong generality across model architectures and downstream tasks.

All papers are available via the provided download links, and the JD team invites readers to discuss the impact of large models on daily life in the comment section.