跳转至

一种计算上可行的因果概率解释框架

文章背景与核心概要

在哲学、科学和政策分析中，解释特定结果发生的原因以及判定各输入因素的责任或功劳至关重要。然而，现有的方法往往走向两个极端：一方面，实际因果性（Actual Causality, AC）理论虽然能给出符合原则的判定，但由于需要穷举反事实场景，除了玩具模型外，在计算上极其昂贵；另一方面，可扩展的归因方法（如标准或因果 SHAP）虽具有较高的计算效率，却常常忽略底层的因果数据生成结构，导致其结论与严谨的因果分析相冲突。

为了弥合这一鸿沟，本文引入了“概率因果影响”（Probabilistic Causal Impact, PCI）框架。PCI 建立在实际因果性以及朱迪亚·珀尔（Pearl）提出的必要性与充分性概率概念基础之上，它将可解释性重新定义为概率因果模型中的一个估计问题，并可通过蒙特卡洛方法进行高效近似。通过定义候选解释分布、反事实值分布以及评分函数，PCI 能够提供易于处理、具因果基础且分级的解释，同时将实际因果性和 Pearl 的因果概率作为其退化情况进行一般化。

摘要 (Summary)

Existing methodologies for explaining specific outcomes and assigning credit or blame are often polarized: * Actual Causality (AC) yields principled verdicts, but is computationally prohibitive for anything beyond toy models due to the requirement of enumerating counterfactual scenarios. * Scalable Attribution Methods (such as standard or causal SHAP) are computationally efficient, but frequently ignore the underlying causal data-generating structure, leading to conclusions that conflict with rigorous causal analysis.

This paper introduces Probabilistic Causal Impact (PCI) to bridge this gap. PCI builds upon actual causality and Pearl’s notions of the probability of necessity and sufficiency, reframing explainability as an estimation problem within a probabilistic causal model that can be efficiently approximated via Monte Carlo methods. By defining a distribution over candidate explanations, a distribution over counterfactual values, and a scoring function, PCI delivers tractable, causally grounded, and graded explanations, generalizing both AC and Pearl's probability of causation as degenerate cases.

现有用于解释特定结果并分配功劳或责任的方法通常呈现两极分化状态： * 实际因果性（AC）能够产生符合原则的判定，但由于需要穷举反事实场景，对于玩具模型之外的任何问题，其计算成本都极其高昂。 * 可扩展的归因方法（如标准或因果 SHAP）虽然计算效率高，但往往忽略了底层生成数据的因果结构，从而导致与严谨的因果分析相冲突的结论。

本文引入了概率因果影响（PCI）来弥合这一鸿沟。PCI 建立在实际因果性以及 Pearl 关于必要性和充分性概率的概念基础之上，将可解释性重新构建为概率因果模型中的估计问题，该问题可以通过蒙特卡洛方法进行高效近似。通过定义候选解释的分布、反事实值的分布以及评分函数，PCI 提供了易于处理、具因果基础且分级的解释，并将 AC 和 Pearl 的因果概率作为退化情况进行了一般化。

论文元数据 (Paper Metadata)

作者 (Authors)

  • Rafal Urbaniak
  • Sam Witty
  • Daniel Waxman
  • Andy Zane
  • Poorva Garg
  • Emily Bunnapradist
  • Sankaran Vaidyanathan
  • Jack Feser
  • Drew Lehe
  • Eli Bingham

摘要原文 (Abstract)

Explaining why a specific outcome occurred, and which inputs deserve the blame or credit, is central to philosophical, scientific, and policy analysis. Existing tools split into two camps. The theory of actual causality (AC) gives principled verdicts, but only for toy-sized models, because computing them requires enumerating counterfactual scenarios. Scalable attribution methods like SHAP (or even causal SHAP) at least partially ignore the causal structure that generated the data, and can give answers that conflict with a careful causal analysis. We close this gap with Probabilistic Causal Impact (PCI).

PCI builds on actual causality and on Pearl's notions of probability of necessity and sufficiency, but recasts the question of explainability as an estimation problem on a probabilistic causal model that is easily approximated via Monte Carlo. By specifying a distribution over "candidate explanations," a distribution over counterfactual values, and a scoring function, PCI provides tractable, causally grounded, graded explanations, generalizing AC and Pearl's probability of causation as degenerate cases.

We evaluate PCI in synthetic and real-world examples, spanning consistency checks with AC, scaling experiments, complex continuous-valued dynamical systems, and a real-world deployed causal machine learning model trained on millions of datapoints.

解释特定结果发生的原因以及哪些输入应承担责任或获得功劳，是哲学、科学和政策分析的核心。现有的工具分为两个阵营。实际因果性（AC）理论给出了符合原则的判定，但仅适用于玩具规模的模型，因为计算它们需要穷举反事实场景。诸如 SHAP（甚至因果 SHAP）等可扩展归因方法至少部分忽略了生成数据的因果结构，并可能给出与仔细的因果分析相冲突的答案。我们通过概率因果影响（PCI）填补了这一空白。

PCI 建立在实际因果性和 Pearl 关于必要性和充分性概率的概念基础之上，但将可解释性问题重新构建为概率因果模型上的估计问题，该模型很容易通过蒙特卡洛进行近似。通过指定“候选解释”的分布、反事实值的分布以及评分函数，PCI 提供了易于处理、具因果基础且分级的解释，将 AC 和 Pearl 的因果概率概括为退化情况。

我们在合成示例和真实世界示例中评估了 PCI，涵盖了与 AC 的一致性检查、规模化实验、复杂的连续值动力系统，以及在数百万个数据点上训练的真实世界部署的因果机器学习模型。

分类详情 (Classification Details)

  • MSC Classes: 68T37, 62D20
  • ACM Classes: I.2.4; I.2.6; G.3