Sequential Conditional Transport on Probabilistic Graphs for Interpretable Counterfactual Fairness

Agathe Fernandes Machado, Arthur Charpentier, Ewen GallicAAAI 2025 · DOI

TL;DR

This paper introduces "sequential conditional transport," a novel method that combines optimal transport with causal graphs to generate counterfactuals for assessing individual fairness. By adapting existing transport techniques to probabilistic graphical models, it offers a more interpretable and computationally efficient way to understand how changing sensitive attributes (like race or gender) would affect an individual's outcome, while keeping other relevant factors constant.

What Problem Does This Paper Address?

Ensuring fairness in AI systems is a critical challenge. While "group fairness" metrics are common, they often fail to capture fairness at the individual level. "Counterfactual fairness" aims to address this by asking: "If this individual's sensitive attribute had been different, would their outcome have changed, all else being equal?" This concept is rooted in causal inference but has been difficult to implement effectively. Existing approaches either rely heavily on causal graphs (which can be complex to build and interpret) or use optimal transport (which can be computationally intensive and less interpretable, especially for non-Gaussian data). This paper bridges this gap by proposing a method that leverages both causal graph structures and optimal transport principles to provide interpretable individual counterfactuals.

Methods

The core contribution is the development of sequential conditional transport. This method builds upon:

  • Optimal Transport (OT): Specifically, it adapts concepts from "Knothe's rearrangement" and "triangular transport." OT provides a mathematical framework for moving probability distributions from one to another, often used to find the "cheapest" way to transform one distribution into another.
  • Causal Graphical Models (DAGs): The approach assumes a known causal graph structure that defines the relationships between variables, including the sensitive attribute (s), legitimate features (x), and the outcome (y).
  • Sequential Transport: Instead of a single, complex multivariate transport map, the method breaks down the transport problem into a sequence of univariate transports, guided by the DAG. This means that for each variable in the causal order, a transport map is computed conditional on the already transported parent variables. This "sequential" nature allows for closed-form solutions and easier interpretation.
  • Conditional Quantiles: The method leverages the connection between conditional quantiles and transport. By mapping conditional cumulative distribution functions (CDFs) to quantile functions, it can derive the transport maps for each variable in the sequence.

This approach is distinct from standard multivariate OT because it exploits the conditional independence structure encoded in the DAG, leading to a more interpretable "fair-twin projection" for counterfactual generation.

Datasets

The paper demonstrates its approach using:

  • Synthetic datasets: Used to illustrate the theoretical foundations and the mechanics of the sequential transport algorithm.
  • German Credit dataset: A real-world dataset used to showcase the application of the method in a practical context.
  • Law School Admission Council dataset: Another real-world dataset used for evaluating counterfactual fairness in logistic regression predictions, comparing the proposed method against existing approaches.

Key Insights and Ideas

  • Interpretable Counterfactuals: By breaking down the transport into sequential, univariate steps guided by a causal graph, the method provides a clear path for how each feature is transformed to create a counterfactual individual. This makes it easier to understand why a counterfactual outcome differs from the factual one.
  • Bridging Causal Graphs and Optimal Transport: The paper successfully integrates the structural insights from causal DAGs with the distributional transformation power of optimal transport, creating a more robust framework for counterfactual fairness.
  • Computational Efficiency for Non-Gaussian Data: Unlike some multivariate OT methods that struggle with non-Gaussian distributions, the sequential approach, by relying on univariate transports and quantile functions, offers closed-form solutions that are more amenable to data-driven estimation and can be applied to new data without re-computation.
  • Individual-Level Fairness Assessment: The primary goal is to move beyond group fairness to assess fairness for each individual, providing a more nuanced understanding of discrimination.

Key Results

  • The proposed sequential conditional transport method is demonstrated to be applicable to both synthetic and real-world datasets.
  • Numerical experiments on the Law School dataset show that the sequential transport method yields results comparable to existing methods like Fairadapt (quantile preservation) and multivariate OT in terms of aggregated counterfactual fairness metrics (e.g., Counterfactual Demographic Parity).
  • The paper highlights that an "aware" model (which explicitly includes sensitive attributes) can sometimes be less counterfactually fair than an "unaware" model, underscoring the importance of counterfactual analysis beyond simply including sensitive attributes.

Limitations and Future Work

  • Causal Graph Assumption: The method relies on a known or correctly specified causal graph. Inferring or learning the causal graph itself is a separate, challenging problem that is not the focus here.
  • Categorical Variables: While the paper mentions handling categorical variables in Appendix B, the core method is primarily described for continuous variables.
  • Future Work: The authors suggest extending this work to:
  • Mitigation strategies: Using sequential transport for pre-processing or in-processing methods to achieve fairness.
  • Handling unknown causal graphs: Incorporating methods for inferring causal structures.
  • More complex causal structures: Extending the approach to handle cycles or latent variables in causal models.

Summary updated

BibTeX
@inproceedings{machado2025sequential,
  author = {Agathe Fernandes Machado and Arthur Charpentier and Ewen Gallic},
  title = {{Sequential Conditional Transport on Probabilistic Graphs for Interpretable Counterfactual Fairness}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i18.34131},
}
Figure