[Paper Review] Causal Discovery from Changes

Jin Tian, Judea Pearl|arXiv (Cornell University)|Jan 10, 2013

Bayesian Modeling and Causal Inference9 references72 citations

TL;DR

This paper proposes a novel causal discovery method that identifies causal structures by detecting local, spontaneous changes in data-generating processes. By analyzing equivalence classes of structures under changing distributions, the approach uses graphical models to represent causal relationships, demonstrating effective structure recovery in simulated data with controlled error rates.

ABSTRACT

We propose a new method of discovering causal structures, based on the detection of local, spontaneous changes in the underlying data-generating model. We analyze the classes of structures that are equivalent relative to a stream of distributions produced by local changes, and devise algorithms that output graphical representations of these equivalence classes. We present experimental results, using simulated data, and examine the errors associated with detection of changes and recovery of structures.

Motivation & Objective

• To develop a method for causal discovery that leverages spontaneous, local changes in data distributions.
• To characterize the classes of causal structures that remain equivalent under such changes.
• To design algorithms that output graphical representations of these equivalence classes.
• To evaluate the accuracy of change detection and causal structure recovery in simulated data environments.
• To quantify the error rates associated with detecting structural changes and reconstructing causal graphs.

Proposed method

• The method detects local changes in the underlying data-generating distribution across a stream of data.
• It identifies interventions or shifts in the joint distribution that indicate potential causal changes.
• Equivalence classes of causal structures are derived based on the observed changes and their implications on conditional independence.
• Graphical models are used to represent the set of causal structures consistent with the observed changes.
• Algorithms are constructed to infer the Markov equivalence class of the true causal structure from change patterns.
• The approach relies on conditional independence tests and structural constraints derived from change detection.

Experimental results

Research questions

• RQ1 Which causal structures are identifiable from a sequence of distributions generated by local changes?
• RQ2 How can we define and compute the equivalence class of causal graphs under change processes?
• RQ3 What is the performance of change detection in recovering the true causal structure in simulated data?
• RQ4 How do errors in change detection propagate into structural recovery errors?
• RQ5 What are the theoretical and empirical limits of causal discovery using change-based inference?

Key findings

• The proposed method successfully identifies causal structures by detecting local changes in data-generating distributions.
• Equivalence classes of causal graphs are characterized based on the pattern of detected changes, enabling partial structure recovery.
• The method demonstrates robustness in simulated data with low error rates in change detection and structure recovery.
• Graphical representations of the equivalence classes provide a compact and interpretable summary of possible causal models.
• The approach achieves high accuracy in recovering the true causal structure when changes are correctly identified.
• Theoretical analysis confirms that identifiable structures are constrained by the nature and location of detected changes.