avatar
朱海
Email:1132863306@qq.com

工作经历

阿里国际 AliExpress · 推荐算法工程师 2024.07 - 至今

  • 专注推荐精排模型商品建模方向，主攻物流偏好建模，深度参与多价模型等工作，覆盖多个核心业务场景，服务线上用户。
  • 主导语义ID在排序模型中的检索与落地，设计并实现GateSID优化方案，提升新商品冷启动曝光率。
  • 参与生成式排序Transformer模型（SORT）的设计与落地，实现订单+6.35%、GMV+5.47%的全场景业务提升。

阿里国际 Lazada · 搜索算法工程师（实习） 2023.07 - 2024.07

  • 针对传统搜索相关性计算仅基于文本信息的局限，负责利用多模态大模型提升搜索相关性，提出Query-LIFE方法。
  • 实现基于查询感知的图文模态融合，有效结合图像和标题信息，增强商品表示准确性。
  • 线上取得订单UV提升3.06%、GMV提升3.19%，成果发表于COLING 2025。

项目经历

语义ID建模（GateSID） 2025.07 - 2026.01

  • 背景：推荐冷启动场景中，排序模型面临协同信号与语言信号权衡难题，新商品因缺少协同信息难以获得足够曝光。
  • 任务：设计能根据商品成熟度动态平衡语言信息与协同信号的推荐框架，提升曝光效率。
  • 行动：提出GateSID框架，利用VQ-VAE将多模态特征离散化为分层语言ID，并构建用户语言ID行为序列；设计门控融合机制，对冷启动商品依据语言信息，对热门商品保留协同信号。
  • 结果：GateSID在所有指标上均优于SOTA基线（COINS、SaviorRec），CTCVR AUC+0.4%、GMV+2.6%、订单量+1.6%，成果投稿SIGIR 2026。成果投稿SIGIR2026

业务级推荐系统排序Transformer模型（SORT） 2025.07 - 至今

  • 背景：单领域数据独立训练导致用户全域意图建模缺失，传统排序模型scaling up收益递减。
  • 任务：设计并落地基于生成式架构的transformer排序大模型，突破传统模型天花板。
  • 行动：实现请求级样本组织、局部注意力；改进transformer架构，引入特殊token、QKNorm、Attention Gate和Sparse MoE模块；优化训练-推理系统，联动工效团队开发注意力算子、算子融合，MFU从3%提升至12%。
  • 结果：线上3万/B测试订单+6.35%、GMV+5.47%；服务效率提升：延迟降低44.67%，吞吐量+121.33%；模型性能：CTR-AUC提升2.1pt，在数据/模型/序列长度维度均展现优异扩展性。详见

搜索相关性多模态（Query-LIFE） 2023.07 - 2024.07

  • 背景：传统相似度计算方法仅基于产品标题和查询建模，无法充分利用图像信息，影响搜索准确度。
  • 任务：设计多模态相似度计算方法，提升搜索结果的准确性和用户体验。
  • 行动：提出Query-LIFE方法，采用基于查询的图文模态融合，有效结合图像和标题信息，利用查询感知的模态对增强商品表示准确度。
  • 结果：线上3万/B测试订单UV提升3.06%、GMV提升3.19%，成果发表于COLING 2025

论文发表

  • Zhu H, Guo Y, Dou R, et al. Query-LIFE: Query-aware Language Image Fusion Embedding for E-Commerce Relevance[C]//Proceedings of the 31st International Conference on Computational Linguistics: Industry Track. 2025: 21-28.
  • Zhu H, Zhao Q, Shang W, et al. Limeattack: Local explainable method for textual hard-label adversarial attack[C]//Proceedings of the AAAI conference on artificial intelligence. 2024, 38(17): 19759-19767.
  • Zhu H, Zhao Q, Wu Y. Beamattack: Generating high-quality textual adversarial examples through beam search and mixed semantic spaces[C]//Pacific-Asia Conference on Knowledge Discovery and Data Mining. Cham: Springer Nature Switzerland, 2023: 454-465.
  • Wang Z, Zheng Y, Zhu H, et al. Transferable adversarial examples can efficiently fool topic models[J]. Computers & Security, 2022, 118: 102749.

教育背景

中国科学技术大学 · 计算机技术（硕士） 2021.09 - 2024.06

  • 省优秀毕业生、中科院院长奖学金、国家奖学金

杭州电子科技大学 · 信息安全（本科） 2017.09 - 2021.06

  • 省优秀毕业生、省政府奖学金、大学生数模竞赛国家二等奖/深圳杯数模国二

Work Experience

AliExpress, Alibaba International · Recommendation Algorithm Engineer 2024.07 - Present

  • Focus on product modeling in recommendation ranking models, specializing in logistics preference modeling, deeply involved in multi-price models, covering multiple core business scenarios, serving online users.
  • Lead the retrieval and implementation of Semantic ID in ranking models, design and implement GateSID optimization scheme, improving exposure rate of new products in cold start.
  • Participate in the design and implementation of generative ranking Transformer model (SORT), achieving +6.35% orders and +5.47% GMV improvement across all scenarios.

Lazada, Alibaba International · Search Algorithm Engineer (Intern) 2023.07 - 2024.07

  • Addressing the limitation of traditional search relevance calculation based only on text information, responsible for leveraging multimodal large models to improve search relevance, proposing Query-LIFE method.
  • Implement query-aware image-text modal fusion, effectively combining image and title information to enhance product representation accuracy.
  • Achieved 3.06% UV order improvement and 3.19% GMV improvement online, results published at COLING 2025.

Project Experience

Semantic ID Modeling (GateSID) 2025.07 - 2026.01

  • Background: In recommendation cold start scenarios, ranking models face the trade-off challenge between collaborative signals and linguistic signals. New products struggle to get sufficient exposure due to lack of collaborative information.
  • Task: Design a recommendation framework that dynamically balances linguistic information and collaborative signals based on product maturity to improve exposure efficiency.
  • Action: Propose GateSID framework, use VQ-VAE to discretize multimodal features into hierarchical language IDs, and build user language ID behavior sequences; design gating fusion mechanism to rely on linguistic information for cold start products and retain collaborative signals for popular products.
  • Result: GateSID outperforms SOTA baselines (COINS, SaviorRec) on all metrics, CTCVR AUC +0.4%, GMV +2.6%, orders +1.6%. Submitted to SIGIR 2026 and KDD 2026. Paper

Business-level Recommendation System Ranking Transformer Model (SORT) 2025.07 - Present

  • Background: Single-domain independent training leads to lack of user global intent modeling, and traditional ranking models show diminishing returns in scaling up.
  • Task: Design and deploy a generative architecture-based transformer ranking large model to break through traditional model ceilings.
  • Action: Implement request-level sample organization and local attention; improve transformer architecture by introducing special tokens, QKNorm, Attention Gate, and Sparse MoE modules; optimize training-inference system, collaborate with efficiency team to develop attention operators and operator fusion, improving MFU from 3% to 12%.
  • Result: Online 30k/B test shows orders +6.35%, GMV +5.47%; Service efficiency: latency reduced by 44.67%, throughput increased by 121.33%; Model performance: CTR-AUC improved by 2.1pt, showing excellent scalability in data/model/sequence length dimensions. Paper

Search Relevance Multimodality (Query-LIFE) 2023.07 - 2024.07

  • Background: Traditional similarity calculation methods model only based on product titles and queries, unable to fully utilize image information, affecting search accuracy.
  • Task: Design multimodal similarity calculation method to improve search result accuracy and user experience.
  • Action: Propose Query-LIFE method, adopt query-based image-text modal fusion, effectively combine image and title information, and use query-aware modal pairs to enhance product representation accuracy.
  • Result: Online 30k/B test shows 3.06% UV order improvement and 3.19% GMV improvement, results published at COLING 2025

Publications

  • Zhu H, Guo Y, Dou R, et al. Query-LIFE: Query-aware Language Image Fusion Embedding for E-Commerce Relevance[C]//Proceedings of the 31st International Conference on Computational Linguistics: Industry Track. 2025: 21-28.
  • Zhu H, Zhao Q, Shang W, et al. Limeattack: Local explainable method for textual hard-label adversarial attack[C]//Proceedings of the AAAI conference on artificial intelligence. 2024, 38(17): 19759-19767.
  • Zhu H, Zhao Q, Wu Y. Beamattack: Generating high-quality textual adversarial examples through beam search and mixed semantic spaces[C]//Pacific-Asia Conference on Knowledge Discovery and Data Mining. Cham: Springer Nature Switzerland, 2023: 454-465.
  • Wang Z, Zheng Y, Zhu H, et al. Transferable adversarial examples can efficiently fool topic models[J]. Computers & Security, 2022, 118: 102749.

Education

University of Science and Technology of China (USTC) · Computer Technology (Master) 2021.09 - 2024.06

  • Provincial Outstanding Graduate, CAS President Scholarship, National Scholarship

Hangzhou Dianzi University · Information Security (Bachelor) 2017.09 - 2021.06

  • Provincial Outstanding Graduate, Provincial Government Scholarship, National Second Prize in Undergraduate Mathematical Modeling Competition/Shenzhen Cup National Second Prize