将相对因果知识付诸实践:基于共享结果私有报告的主干可识别性
文章背景与核心概要
在多智能体系统与分布式决策中,不同智能体往往拥有各自独立的结构因果模型(SCMs),如何有效交换因果知识一直是该领域的核心挑战。“因果知识相对性(Relativity of Causal Knowledge, RCK)”框架指出,智能体之间可以通过一个被称为“主干(backbone)”的共享、干预一致性抽象来进行知识交流。然而,一个根本性的识别问题随之产生:在什么条件下,才能仅从各个智能体的私有因果报告中唯一确定这个主干?
本文针对这一问题展开了深入研究。作者证明了在共同效应场景(即两个私有原因影响一个共享结果)下,局部的因果边际分布并不足以唯一识别该主干,并数学上证明了存在无限多个联合干预核(intervention kernels),它们在产生相同私有报告的同时,对联合干预的预测却并不一致。为了解决这一难题,作者提出了一种通过传递“经因果识别的响应函数”来实现识别的方案,并以教育增值(value-added)评估为例进行了生动阐述,将该挑战本质上界定为一个在政策合成之前的通信问题。
摘要 (Summary)
The Relativity of Causal Knowledge (RCK) framework describes how agents with distinct structural causal models can exchange knowledge via a shared, interventionally consistent abstraction known as a "backbone." This paper addresses the fundamental identification problem: under what conditions can this backbone be determined from the private causal reports of individual agents?
“因果知识相对性(Relativity of Causal Knowledge, RCK)”框架描述了拥有不同结构因果模型的智能体如何通过一个被称为“主干”的共享、干预一致性抽象来交换知识。本文探讨了其根本的识别问题:在什么条件下,可以从各个智能体的私有因果报告中确定这个主干?
The authors demonstrate that in a common-effect scenario—where two private causes influence a single shared outcome—local causal marginals are insufficient to uniquely identify the backbone. They prove that infinitely many joint intervention kernels can produce identical private reports while disagreeing on joint interventions. The paper concludes by proposing a solution through the communication of causally identified response functions and illustrates this with an education value-added example, framing the challenge as a communication problem before it becomes a policy-composition problem.
作者证明了在共同效应场景(即两个私有原因影响一个单一的共享结果)中,局部的因果边际分布不足以唯一识别主干。他们证明了存在无限多个联合干预核,它们可以产生相同的私有报告,但在联合干预上却存在分歧。最后,本文提出了一种通过通信传输经因果识别的响应函数的解决方案,并以教育增值评估为例进行了说明,将这一挑战在演变为政策合成问题之前,先定位为一个通信问题。
核心贡献 (Key Contributions)
- Identifiability Analysis: Demonstrates that under standard assumptions (compatibility, non-degeneracy, and local overlap), local causal marginals fail to uniquely identify the joint backbone.
- Mathematical Proof: Shows that there exist infinitely many joint intervention kernels consistent with the same private reports.
- Conditional Recovery: Establishes that additive separability removes certain degrees of freedom, though observational residual summaries remain insufficient for full identification.
- Communication Framework: Proposes that identification is achievable when agents communicate causally identified response functions, shifting the focus from mere data sharing to structured causal communication.
- 可识别性分析: 证明了在标准假设(相容性、非退化性和局部重叠)下,局部因果边际分布无法唯一识别联合主干。
- 数学证明: 表明存在无限多个与相同私有报告一致的联合干预核。
- 条件恢复: 确立了加性可分性消除了某些自由度,尽管观测残差摘要对于完全识别仍然是不够的。
- 通信框架: 提出当智能体交流经因果识别的响应函数时,识别是可行的,从而将重心从单纯的数据共享转移到结构化的因果通信上。
获取与资源 (Access & Resources)
- 查看 PDF
- HTML(实验性)
- TeX 源码
- 许可协议: 知识共享署名 4.0 国际版

提交历史 (Submission History)
- [v1] Tue, 11 Aug 2026 08:46:18 UTC
- [v1] 2026年8月11日 周二 08:46:18 UTC