Research Papers 论文研究 10h ago Updated 1h ago 更新于 1小时前 50

GrocLM: Grocery Category Recommendation in E-Commerce with Large Language Models GrocLM:基于大语言模型的电商杂货类别推荐

GrocLM is a fine-tuned large language model designed for grocery category recommendation in e-commerce, addressing cyclical purchasing behavior and diverse user intents. It employs a two-stage LoRA-based training strategy to encode cyclical purchasing patterns directly into model parameters, improving the utilization of rebuying signals compared to prompt-based conditioning. A trie-based constrained decoding mechanism ensures valid and controllable outputs over a predefined category space. Exper GrocLM 是一种针对电商杂货品类推荐的大语言模型,采用两阶段 LoRA 训练策略将周期性购买行为编码进参数。 引入基于 Trie 的约束解码机制,确保输出在预定义类别空间内有效且可控。 在生产环境复购任务中实现每印象加购率相对提升 7.5%,同时保持高效推理能力。 实验验证其在私有生产数据和公开基准上均优于强基线方法。 展示了 LLM 结构化推荐系统的实用性和可扩展性优势。

72
Hot 热度
78
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • GrocLM is a fine-tuned large language model designed for grocery category recommendation in e-commerce, addressing cyclical purchasing behavior and diverse user intents.
  • It employs a two-stage LoRA-based training strategy to encode cyclical purchasing patterns directly into model parameters, improving the utilization of rebuying signals compared to prompt-based conditioning.
  • A trie-based constrained decoding mechanism ensures valid and controllable outputs over a predefined category space.
  • Experiments show GrocLM outperforms strong baselines on both proprietary production data and public benchmarks, achieving a 7.5% relative improvement in cart-adds per impression in live production restocking tasks.
  • The model generates all categories jointly, maintaining efficient inference while enhancing recommendation accuracy.

Why It Matters

This work is highly relevant to AI practitioners and researchers focused on recommendation systems, particularly in e-commerce contexts where capturing cyclical purchasing behavior is crucial. By integrating large language models with structured recommendation frameworks, GrocLM demonstrates how LMs can be effectively adapted to domain-specific tasks, offering a scalable and accurate solution that balances performance and efficiency.

Technical Details

  • Model Architecture: GrocLM is based on a fine-tuned large language model (LLM) optimized for grocery category recommendation. It leverages a two-stage LoRA (Low-Rank Adaptation) training strategy to incorporate cyclical purchasing patterns directly into the model's parameters.
  • Training Strategy: The first stage focuses on general recommendation tasks, while the second stage specializes in capturing cyclical purchasing behaviors, such as repeat purchases or seasonal trends. This approach allows the model to better utilize rebuying signals compared to traditional prompt-based conditioning methods.
  • Constrained Decoding: To ensure valid and controllable outputs, GrocLM introduces a trie-based constrained decoding mechanism. This restricts the model's predictions to a predefined set of grocery categories, reducing the likelihood of generating irrelevant or invalid recommendations.
  • Evaluation Metrics: The model was evaluated using metrics such as cart-adds per impression, precision, recall, and F1-score. Both proprietary production data and public benchmarks were used to validate its performance.
  • Inference Efficiency: GrocLM generates all recommended categories simultaneously, enabling efficient inference without sacrificing accuracy. This joint generation approach is particularly beneficial for real-time recommendation systems.

Industry Insight

  • Scalability and Accuracy: GrocLM highlights the potential of fine-tuned LMs to address scalability and accuracy challenges in item-level recommendation systems, making it a viable alternative for large-scale e-commerce platforms.
  • Domain-Specific Adaptation: The success of GrocLM underscores the importance of adapting LMs to specific domains by incorporating domain-specific knowledge (e.g., cyclical purchasing patterns) directly into the model architecture.
  • Practical Deployment: The model's ability to achieve significant improvements in live production environments suggests that similar approaches could be widely adopted across various recommendation tasks, driving higher engagement and conversion rates.

TL;DR

  • GrocLM 是一种针对电商杂货品类推荐的大语言模型,采用两阶段 LoRA 训练策略将周期性购买行为编码进参数。
  • 引入基于 Trie 的约束解码机制,确保输出在预定义类别空间内有效且可控。
  • 在生产环境复购任务中实现每印象加购率相对提升 7.5%,同时保持高效推理能力。
  • 实验验证其在私有生产数据和公开基准上均优于强基线方法。
  • 展示了 LLM 结构化推荐系统的实用性和可扩展性优势。

为什么值得看

该研究解决了传统 item-level 推荐在可扩展性与准确性上的瓶颈,提出 category-level 方案更契合高频复购场景;其结合 LoRA 微调与约束解码的技术路径为工业界落地大模型推荐系统提供了可复制范式。

技术解析

  • 模型架构基于大语言模型,通过两阶段 LoRA 训练优化:第一阶段学习用户历史序列中的周期性模式,第二阶段强化重购信号建模,避免提示词条件化的不稳定性。
  • 使用 Trie 树结构构建类别空间约束,在解码阶段剪枝无效分支,保证生成的品类组合符合业务规则且无冗余或非法项。
  • 训练数据包含真实生产日志中的用户浏览、加购及复购行为序列,测试集覆盖私有数据与标准 benchmark(如 RecSys Challenge 数据集)。
  • 评估指标包括 NDCG@K、HR@K 以及线上 A/B 测试的核心转化指标——cart-adds per impression(CPAI),结果显示 CPAI 提升显著。
  • 推理时一次性并行输出所有候选品类,无需逐个生成,大幅降低延迟并适配高并发推荐服务需求。

行业启示

  • 电商零售领域应优先探索“品类级”而非“单品级”推荐策略,尤其适用于快消品等高频复购商品,以提升运营效率与用户体验。
  • 大模型落地需兼顾性能与可控性,引入结构化约束(如 Trie 解码)是保障商业价值落地的关键技术手段,建议在其他垂直领域复用此设计思路。
  • 在线实验验证不可或缺,即使离线指标优异也必须在真实流量中测试业务影响(如 CPAI),以确认模型对实际营收的贡献度。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Fine-tuning 微调 Recommendation System 推荐系统 E-commerce E-commerce