文章背景与核心概要
随着服务型企业向动态打包产品转型，传统的基于树模型的分类器在处理稀疏、不平衡数据以及实时会话转录等多模态信号时显得力不从心。本文详细记录了一个生产级对话推荐系统从传统的梯度提升树（CatBoost）模型向深度学习架构转型的全过程。
作者成功实现了一种成对二分类（pairwise-binary）深度推荐器，它在保持高性价比性能的同时，支持客服坐席进行实时的“动态推销（dynamic pitching）”。该研究不仅评估了双塔模型与 DeepFM 等深度架构，还创新性地引入了注意力池化机制来处理长对话上下文，并通过噪声注入和负采样提升了模型的鲁棒性。
从梯度提升树到深度推荐器：生产环境客服推荐系统迁移的实战经验
作者： Sonia Sharma, Jeyendran Balakrishnan, Shreya Rajpal, Swapnil Parekh, Nagaraj Janardhana, Andrew Mattarella-Micke
日期： 2026年8月25日
arXiv ID： 2608.24132
学科： 机器学习 (cs.LG)；人工智能 (cs.AI)
摘要
本文详细介绍了生产级对话推荐系统从传统的梯度提升树（CatBoost）模型向基于深度学习的架构过渡的过程。随着服务企业向动态、打包的服务产品转型，传统的基于树的分类器难以应对稀疏、不平衡的数据，且无法处理诸如实时会话转录等多模态信号。作者成功实现了一种成对二分类（pairwise-binary）深度推荐器，在保持高性能的同时，使客服代表能够进行实时的“动态推销”。
This paper details the transition of a production-grade conversational recommendation system from a traditional gradient-boosted tree (CatBoost) model to a deep learning-based architecture. As service businesses move toward dynamic, bundled offerings, traditional tree-based classifiers struggle with sparse, imbalanced data and the inability to process multimodal signals like live conversation transcripts. The authors successfully implemented a pairwise-binary deep recommender that maintains high performance while enabling real-time "dynamic pitching" for support agents.
核心技术贡献
1. 架构转变
- 从多分类到成对二分类： 团队将推荐任务重新构想为成对二分类预测问题。这使得模型能够联合学习用户和物品特征，比树模型固定的标签空间提供了更大的灵活性。
- 模型架构： 该研究评估了几种深度学习架构，包括双塔模型和 DeepFM 变体，并利用对比损失等损失函数来提高推荐准确率。
1. Architectural Shift
- From Multiclass to Pairwise-Binary: The team reformulated the recommendation task as a pairwise-binary prediction problem. This allows the model to learn jointly from user and item features, providing greater flexibility than the fixed label space of tree-based models.
- Model Architectures: The study evaluates several deep learning architectures, including two-tower models and DeepFM variants, utilizing loss functions such as contrastive loss to improve recommendation accuracy.
2. 处理会话上下文
- 注意力池化： 为了处理漫长、实时的会话转录，作者在转录文本块（chunks）上实现了注意力池化（attention pooling）。
- 基准测试： 新方法与传统的 TF-IDF 和句子嵌入（sentence-embedding）基线进行了对比，在捕捉相关上下文方面表现出更优越的性能。
2. Handling Conversational Context
- Attention Pooling: To process long, live conversation transcripts, the authors implemented attention pooling over transcript chunks.
- Benchmarking: The new approach was benchmarked against traditional TF-IDF and sentence-embedding baselines, demonstrating superior performance in capturing relevant context.
3. 特征增强
- 噪声注入与负采样： 为了提高所学表征的鲁棒性，作者采用了系统的负采样和噪声注入技术。
3. Feature Enhancement
- Noise Injection & Negative Sampling: To improve the robustness of learned representations, the authors employed systematic negative sampling and noise injection techniques.
性能表现
此次迁移受制于一个不可妥协的约束：维持实时的推荐质量。结果表明： * 初期性能对等： 在对话刚开始时，深度学习方法达到了与 CatBoost 基线相当的性能。 * 后期表现优异： 在对话的后期阶段，新模型显著优于传统的树基系统，证明了深度学习在处理不断演变的、多轮对话上下文方面的价值。
Performance Outcomes
The migration was governed by a non-negotiable constraint: maintaining live recommendation quality. The results show: * Parity at Start: The deep learning approach achieves performance parity with the CatBoost baseline at the beginning of conversations. * Outperformance: The new model significantly outperforms the legacy tree-based system during later stages of the conversation, proving the value of deep learning in handling evolving, multi-turn dialogue context.
访问论文
Accessing the Paper