关系型结构因果模型

文章背景与核心概要

现代人工智能系统需要构建既具备因果性（支持对干预和反事实的推理）又具备组合性（支持泛化到未见过的对象组合）的环境模型。本文引入了关系型结构因果模型（Relational Structural Causal Models），将传统的结构因果模型扩展到了对象和关系动态变化的环境中。

该研究的主要贡献包括：证明了在没有结构假设的情况下，无法识别对未见对象组合的观测和因果查询；引入了关系因果图和符号识别准则以处理可识别性问题（甚至在存在未观测混杂因子的情形下）；并通过模拟包含汽车、信号灯和行人的动态交通场景，证明了关系型神经网络因果模型的有效性，其性能显著优于非关系型的基线方法。

摘要 (Summary)

现代人工智能需要构建具备环境模型的系统，该模型既要具有因果性（支持对干预和反事实的推理），又要具有组合性（支持泛化到未见过的对象组合）。本文正式研究了何时以及如何学习此类模型。我们开发了关系型结构因果模型，将结构因果模型（Pearl 2009）扩展到对象及其关系发生变化的场景中。首先，我们证明了在没有进一步假设的情况下，不仅是因果查询，连关于未见对象组合的观测查询都无法被识别。为了实现这种识别——包括在存在未观测混杂因子的空中——我们定义了关系因果图并推导出了符号识别准则。最后，我们提出了关系型神经因果模型，这是一种经证明正确的方法，在包含各种汽车、信号灯和行人的模拟交通场景中，其表现优于非关系型基线。

Modern artificial intelligence requires environmental models that are both causal (supporting reasoning about interventions and counterfactuals) and combinatorial (supporting generalization to unseen combinations of objects). This paper introduces Relational Structural Causal Models, extending traditional structural causal models to dynamic environments where objects and relations vary.

Key contributions include: * Proving that observational and causal queries over unseen object combinations cannot be identified without structural assumptions. * Introducing relational causal graphs and symbolic identification criteria to handle identification (even in the presence of unobserved confounding). * Proving the efficacy of relational neural causal models, which outperform non-relational baselines on simulated traffic scenes involving dynamic combinations of cars, signals, and pedestrians.

文档元数据 (Document Metadata)

Field	Details
arXiv ID	arXiv:2606.14892 [cs.AI]
Authors	Adiba Ejaz, Elias Bareinboim
Primary Subject	Artificial Intelligence (cs.AI)
Other Subjects	Machine Learning (cs.LG), Social and Information Networks (cs.SI), Machine Learning (stat.ML)
Submitted	June 12, 2026 (Last revised August 22, 2026, v2)
Conference	Proceedings of the Forty-Third International Conference on Machine Learning (ICML)

摘要详情 (Abstract)

人工神经网络必须具备其环境的模型，该模型既是因果的（支持关于干预和反事实的推理），也是组合的（支持对未见过的对象组合进行泛化）。在这项工作中，我们正式研究了何时以及如何学习这样一种模型。我们开发了关系型结构因果模型，将结构因果模型（Pearl 2009）扩展到对象及其关系变化的设置中。首先，我们展示了如果没有进一步的假设，不仅是因果查询，连关于未见对象组合的观测查询都无法被识别。为了实现这种识别——包括在存在未观测混杂因素的情况下——我们定义了关系因果图并推导出了符号识别标准。最后，我们提出了关系型神经因果模型，这是一种可证明正确的方法，在具有变化的汽车、信号和行人的模拟交通场景中，其性能优于非关系型基线。

An artificial intelligence must have a model of its environment that is causal, supporting reasoning about interventions and counterfactuals, and also combinatorial, supporting generalization to unseen combinations of objects. In this work, we formally study when and how such a model can be learned. We develop relational structural causal models, extending structural causal models (Pearl 2009) to settings where objects and their relations vary. First, we show how answers to not only causal but also observational queries about unseen combinations of objects can not be identified without further assumptions. To enable such identification—including in the presence of unobserved confounding—we define relational causal graphs and derive symbolic identification criteria. Finally, we propose relational neural causal models, a provably correct approach that outperforms non-relational baselines on simulated traffic scenes with varying cars, signals, and pedestrians.

资源与链接 (Resources & Links)

• 全文访问： 查看 PDF | HTML（实验性） | TeX 源码
• 代码与实现： GitHub 仓库
• 数字对象唯一标识符 (DOI)： 10.48550/arXiv.2606.14892

• Full-Text Access: View PDF | HTML (Experimental) | TeX Source
• Code & Implementation: GitHub Repository
• Digital Object Identifier (DOI): 10.48550/arXiv.2606.14892

文章许可协议 (Article License)

知识共享署名 4.0 国际许可协议 (Creative Commons Attribution 4.0 International)

Creative Commons Attribution 4.0 International