文献阅读 250102-Decomposing causality into its synergistic, unique, and redundant components

最新推荐文章于 2025-03-07 16:30:53 发布

ZzYH22

最新推荐文章于 2025-03-07 16:30:53 发布

阅读量1.1k

点赞数 20

文章标签：笔记

本文链接：https://blog.csdn.net/weixin_46297571/article/details/144884959

版权

Decomposing causality into its synergistic, unique, and redundant components

来自 <Decomposing causality into its synergistic, unique, and redundant components | Nature Communications>

将因果关系分解为其协同、独特和冗余的组成部分

来自 <Decomposing causality into its synergistic, unique, and redundant components | Nature Communications>

Code for this word: https://github.com/Computational-Turbulence-Group/SURD

该方法结合了因果关系泄漏的概念，它通过评估未观察到的变量的影响来充当质量控制机制。这种衡量标准减轻了假设因果充分性的需求（即，必须准确衡量变量的所有常见原因），因为它提供了一个可量化的衡量标准，即来自未观察到的变量的信息在多大程度上仍未被解释。该方法无需模型，即不需要有关系统动力学的先验知识。这使得 SURD 对涉及具有线性和非线性依赖关系的确定性或随机多变量系统的应用程序很有吸引力。

## Key Pts:

Current methods for causal inference face significant challenges due to nonlinear dependencies, stochastic interactions, self-causation, collider effects (碰撞效应), and influences from exogenous factors, among others.
Existing methods can address some of these, but none single approach can integrated all these aspects. Here the authors arise a method named "SURD" (Synergistic-Unique-Redundant Decomposition of causality) to quantify causality as the increments of redundant, unique, and synergistic(冗余、独特和协同) information gained about future events from past obs. .

## Intro:

A central aspect of causality is the concept of physical influence: manipulation of the cause manifests as changes in the effects. Eg: exposure to air pollution has a causal connection to a higher incidence of chronic respiratory.
The precise definition of causality remains elusive, yet it must be distinguished from the concepts of association and correlation.

Association indicates a statistical relationship between two variables in which they have a tendency to co-occur more often than would be expected by random chance. Yet, association does not automatically imply causation15. Association may arise from shared causes, statistical coincidences, or the influence of confounding factors.
Correlation refers to a particular type of association that measures the monotonic strength and direction of variables. Correlation implies association but not causation; causation implies association but not correlation.

## Methods:

Mediator variables (A → B → C) : In this causal chain, B serves as a mechanism or mediator responsible for transmitting the influence of A to C. "Mediator variables help explain the underlying mechanisms by which an independent variable influences a dependent variable. "
Confounding variables (A ← B → C) : B acts as the common cause of two variables. Consider this structure, B may create a statistical correlation between A and C, even if there is no direct causal link between them. "Confounding variables can obscure or distort the genuine relationship between variables. "
Collider variables (A → B ← C): This causal structure represent the effect of multiple factors acting on the same variable. In nonlinear dynamical systems, most variables are affected by multiple causes due to their coupling. A collider exhibits redundant causes when both A and C contribute to the same effect or outcome of B, creating overlapping or duplicative influences on the outcome.

Traditional methods:

The approach offers a pathway for evaluating the causal effect that the process A exerts on another process B by setting A at a modified value A- and observing the consequences of the post-intervention in B. Based on the intervention, Granger causality (GC) measures the causality from the process B to A by evaluating how the inclusion of B in an autoregressive model reduces the forecast error for A.

…

Peter-Clark algorithm (PC), with subsequent extensions incorporating tests for momentary conditional independence (PCMCI). PCMCI aims to optimally identify a reduced conditioning set that includes the parents of the target variable.

Vector Q = [Q1(t), Q2(t),…, QN(t)] is the collection of N time-dependent vars. The components are observables and treated as r.v.

Objective is to quantify the causality from the components of Q to the future of the target var Qj, denoted Qj+ = Qj(t + delta T), delta T >0.

SURD 将因果效应量化为从观测的福利成员或成员组获取的Qj+的增加信息 delta I.

Qj+的信息由Shannon 熵量化，定义为H(Qj+)，代表明确确定Qj+所需要的平均bit数。

高度不确定（高熵）的过程也是我们在确定其状态时获得最多信息的过程。相反，当过程完全确定时，不确定性为零，表示当结果被揭示时没有获得任何信息。利用信息向前传播的原则：