跳转至

文章背景与核心概要

统一推荐模型通常需要结合非序列多域特征和序列用户行为，然而传统的以交互为中心的方法主要集中在单个层内部混合异构 Token。本文指出，扩展统一推荐模型还需要积极管理意图信息在堆叠层之间的携带、过滤和保留方式。

受基于流的表示动力学的启发，作者引入了特征传输（feature transport）这一范式，将深层统一推荐视为一个离散的、受上下文条件约束的表示演化过程。他们提出了 CRAFT（Contextual Residual Adaptive Feature Transport，上下文残差自适应特征传输）这一新型构建模块，它将非序列特征概括为一个感知可靠性的上下文域，以此来指导序列表示的残差位移和记忆保留信号。该研究在 TAAC 2026 广告推荐比赛中取得了优异成绩，并通过多层堆叠与维度扩展验证了其强大的可扩展性和泛化能力。

From Feature Interaction to Feature Transport: A Unified Block for Scalable Recommendation Models

From Feature Interaction to Feature Transport: A Unified Block for Scalable Recommendation Models

arXiv ID: arXiv:2609.01655
Subjects: Information Retrieval (cs.IR); Artificial Intelligence (cs.AI)
Authors: Zichen Luo, Jiachen Guo, Keming Gu, Jie Zhang
Submitted: August 31, 2026 (Accepted to KDDCUP 2026 Workshop)
Source Code: GitHub - CRAFT

arXiv ID: arXiv:2609.01655
Subjects: Information Retrieval (cs.IR); Artificial Intelligence (cs.AI)
Authors: Zichen Luo, Jiachen Guo, Keming Gu, Jie Zhang
Submitted: August 31, 2026 (Accepted to KDDCUP 2026 Workshop)
Source Code: GitHub - CRAFT

📌 Summary

📌 Summary

统一推荐模型通常结合了非序列多域特征和序列用户行为，但传统的侧重交互的方法大多集中在单个层内混合异构 Token。本文认为，扩展统一推荐模型还需要主动管理意图信息在堆叠层之间的携带、过滤和保留方式。

Unified recommendation models typically combine non-sequential multi-field features and sequential user behaviors, but conventional interaction-focused methods mostly concentrate on mixing heterogeneous tokens within individual layers. This paper argues that scaling unified recommenders also requires actively managing how intent information is carried, filtered, and preserved across stacked layers.

借鉴基于流的表示动力学的灵感，作者引入了特征传输（feature transport）——一种将深度统一推荐视为离散的、受上下文条件约束的表示演化过程的范式。他们提出了 CRAFT（Contextual Residual Adaptive Feature Transport，上下文残差自适应特征传输），这是一个新颖的构建模块，它将非序列特征概括为一个感知可靠性的上下文域，从而为序列表示指导残差位移和记忆保留信号。

Drawing inspiration from flow-based representation dynamics, the authors introduce feature transport—a paradigm that treats deep unified recommendation as a discrete, context-conditioned representation evolution process. They propose CRAFT (Contextual Residual Adaptive Feature Transport), a novel block that summarizes non-sequential features into a reliability-aware contextual field to guide residual displacement and memory-preserving signals for sequence representations.

核心亮点：

  • 主动控制： 非序列上下文充当表示演化的主动控制器，而非被动的交互目标。
  • 比赛成功： 在 TAAC 2026 广告推荐比赛中，CRAFT 取得了测试集 AUC 0.838090 的成绩，超越了此前排行榜最优的 0.83798。
  • 可扩展性： 将 CRAFT 堆叠至六个模块可进一步将测试集 AUC 提升至 0.838148，而扩大隐藏维度则达到了 0.838106，展现出强大的可扩展性和泛化能力。

Key Highlights:

  • Active Control: Non-sequential context acts as an active controller for representation evolution rather than a passive target for interaction.
  • Competition Success: In the TAAC 2026 advertising recommendation competition, CRAFT achieved a test AUC of 0.838090, outperforming the previous leaderboard-best score of 0.83798.
  • Scalability: Stacking CRAFT to six blocks further raises the test AUC to 0.838148, while expanding the hidden dimension achieves 0.838106, demonstrating robust scalability and generalization.