Optimal transport and Wasserstein distances for causal models

Patrick Cheridito, Stephan EcksteinBernoulli 2023 · arXiv

TL;DR

This paper introduces a novel framework for optimal transport that explicitly incorporates the causal structure of directed acyclic graphs (DAGs). By defining "G-causal couplings" and corresponding "G-causal Wasserstein distances," the authors provide a more nuanced way to compare probability distributions that respects causal relationships, leading to improved continuity of causal effect estimates and more robust interpolations.

What Problem Does This Paper Address?

Traditional optimal transport and Wasserstein distances are powerful tools for comparing probability distributions, but they often treat variables as independent or ignore their underlying causal dependencies. In fields like causal inference, understanding how changes in one variable propagate through a system is crucial. Standard Wasserstein distances can be insensitive to these causal structures, leading to issues like the average treatment effect not being continuous with respect to small changes in the underlying probability model. This paper aims to bridge this gap by developing optimal transport methods that are "causally aware," ensuring that the transport plans and distances reflect the directed dependencies defined by a causal graph.

Methods

The core contribution is the definition of G-causal couplings and G-causal transport plans, where 'G' refers to a directed graph representing causal relationships. A coupling is G-causal if it respects the structure of G, meaning the way variables are related in the coupling mirrors the causal dependencies. This is formalized by requiring that the conditional distribution of a variable depends only on its direct causes (parents in the graph) and potentially other variables that are themselves determined by those causes.

Based on these G-causal couplings, the authors define G-causal Wasserstein distances. These distances are essentially optimal transport costs, but the optimization is restricted to G-causal couplings. This restriction leads to a finer topology on the space of probability measures, which is more sensitive to causal structure. The paper also explores G-bicausal couplings, which are G-causal in both directions, and investigates their properties for directed acyclic graphs (DAGs).

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Causality Matters for Transport: The paper demonstrates that ignoring causal structure in optimal transport can lead to a loss of sensitivity to important relationships, particularly for estimating causal effects.
  • G-causal Distances are Finer: By restricting transport plans to those respecting causal graphs, the resulting G-causal Wasserstein distances provide a more refined metric for comparing distributions, especially when causal relationships are complex.
  • Improved Continuity of Causal Effects: A key finding is that average treatment effects, which are often not continuous with respect to standard Wasserstein distances, become Lipschitz continuous with respect to G-causal Wasserstein distances. This is crucial for robust causal inference and error estimation.
  • G-compatible Interpolations: The paper introduces a notion of G-causal interpolation that, under certain conditions, preserves the causal compatibility of the interpolated distributions, unlike standard Wasserstein interpolations.

Key Results

  • The paper provides formal definitions of G-causal couplings and G-causal Wasserstein distances.
  • It establishes equivalences between different characterizations of G-causal couplings for DAGs.
  • It proves that average treatment effects are Lipschitz continuous with respect to G-causal Wasserstein distance (WG,1) under suitable assumptions, while not being continuous with respect to standard Wasserstein distance (W1).
  • It shows that small perturbations in structural causal models lead to small deviations in G-causal Wasserstein distance.
  • It introduces G-causal interpolations and provides conditions under which they preserve G-compatibility.
  • A counterexample is provided to show that WG,1, in general, does not satisfy the triangle inequality, meaning it's a semimetric.

Limitations and Future Work

  • Triangle Inequality: The G-causal Wasserstein distances are shown to be semimetrics, not full metrics, as they do not always satisfy the triangle inequality. This is a limitation for some theoretical applications.
  • Computational Aspects: While the paper defines the concepts, the computational efficiency of calculating these G-causal Wasserstein distances for general graphs is not deeply explored.
  • General Graphs: The paper focuses heavily on DAGs. Extending the framework to graphs with cycles might be a future direction.
  • Practical Applications: The paper highlights the theoretical benefits for causal inference. Future work could explore concrete applications in areas like policy evaluation or fairness in machine learning where causal relationships are paramount.

Summary updated

Abstract

In this paper, we introduce a variant of optimal transport adapted to the causal structure given by an underlying directed graph $G$. Different graph structures lead to different specifications of the optimal transport problem. For instance, a fully connected graph yields standard optimal transport, a linear graph structure corresponds to causal optimal transport between the distributions of two discrete-time stochastic processes, and an empty graph leads to a notion of optimal transport related to CO-OT, Gromov-Wasserstein distances and factored OT. We derive different characterizations of $G$-causal transport plans and introduce Wasserstein distances between causal models that respect the underlying graph structure. We show that average treatment effects are continuous with respect to $G$-causal Wasserstein distances and small perturbations of structural causal models lead to small deviations in $G$-causal Wasserstein distance. We also introduce an interpolation between causal models based on $G$-causal Wasserstein distance and compare it to standard Wasserstein interpolation.
BibTeX
@inproceedings{cheridito2023optimal,
  author = {Patrick Cheridito and Stephan Eckstein},
  title = {{Optimal transport and Wasserstein distances for causal models}},
  booktitle = {Bernoulli},
  year = {2023},
}
Figure