Counterfactual Distribution Regression for Structured Inference

Nicolò Colombo, Ricardo Silva, Soong Moon Kang, Arthur GrettonarXiv 2019 · arXiv

Introduction

Predicting how complex, interconnected systems react to unexpected disruptions is a fundamental challenge in urban planning and policy management. When a segment of a subway network closes, the impact is rarely localized; it ripples through the entire system as passengers reroute or change their destinations. Traditional causal inference typically focuses on estimating the "average treatment effect," such as the average increase in traffic across a city. However, effective management often requires understanding the full distribution of outcomes to account for variability and extreme events.

This paper introduces a novel framework for "structured inference," addressing the problem of how to use historical data from a system's natural state and a limited set of past disruptions to predict the impact of a novel perturbation at a previously unaffected location.

Methodology: Distribution-to-Distribution Regression

The authors frame the prediction task as a distribution-to-distribution regression problem. Rather than mapping individual data points, the model learns a mapping between entire probability distributions. This approach relies on several key technical components:

1. Counterfactual Covariates

The model utilizes the "natural regime" distribution—the system's state during normal operations—as a primary covariate. During a disruption, this distribution serves as a counterfactual, providing a probabilistic baseline of what would have occurred had the shock not taken place.

2. RKHS Mean Embeddings

To make these complex distributions mathematically tractable, the authors embed them into a Reproducing Kernel Hilbert Space (RKHS). This process transforms a probability distribution into a "mean embedding," effectively a single point in a high-dimensional feature space where linear operations can be performed.

3. Linear Operators

The relationship between the natural distribution and the perturbed distribution is modeled as a linear operator ($L_\Psi$) acting on these embeddings. This allows the model to capture the transformation from normal to disrupted states without explicitly modeling every individual agent's decision-making process.

4. Efficient Sampling

A significant technical contribution is a new "mean embedding sampling" scheme. While traditional methods like "herding" are computationally expensive, the authors propose a simplex-constrained convex minimization approach to reconstruct samples from the predicted distribution more efficiently, facilitating real-time application.

Experimental Setup: The London Underground

The framework was evaluated using a real-world dataset from the London Underground, consisting of a graph of 269 stations and 11 lines. The data included:

  • Natural Regime: 35 days of journey records (entry/exit times and locations) from late 2013.
  • Disruptions: A database of 72 observed disruptions, including start/end times and the specific Region of Interest (ROI) affected.
  • Evaluation: The study focused on the 20 disruptions with the highest "observable score," representing the most significant deviations from normal traffic patterns.

Key Insights and Results

The Power of the Natural State

A primary insight of the research is that a system's behavior during a crisis is highly predictable if its behavior during normal operations and its physical constraints (the network graph) are known. By incorporating the graph structure (adjacency matrix) and spatial interference, the model assumes that the effects of a perturbation decay or shift based on network distance.

Performance Metrics

The Counterfactual Distribution Regression model was compared against baseline natural regime models and random models:

  • Log-Likelihood: The proposed model significantly outperformed the baseline in predicting exit counts during disruption days, successfully capturing how passenger distributions shift and peak during shocks.
  • Error Reduction: Quantitative evaluations showed a marked reduction in the logarithm of the square error compared to both baseline and random models.
  • Qualitative Accuracy: Visualizations of station-specific densities demonstrated that the model correctly identifies when traffic will "spike" at alternative stations—shifts that the baseline model failed to predict.

Conclusion and Future Work

By sidestepping the need to model latent variables or individual passenger choices, this framework provides a robust method for aggregate system prediction. However, the model currently assumes that all perturbations belong to a similar class (e.g., unplanned closures) and relies on a static network graph. Future research may explore dynamic network structures or scenarios where the nature of the shock is fundamentally different from historical data, such as city-wide strikes. Despite these limitations, the approach offers a powerful tool for understanding and mitigating the impacts of disruptions in complex infrastructure.

Summary updated

Abstract

We consider problems in which a system receives external \emph{perturbations} from time to time. For instance, the system can be a train network in which particular lines are repeatedly disrupted without warning, having an effect on passenger behavior. The goal is to predict changes in the behavior of the system at particular points of interest, such as passenger traffic around stations at the affected rails. We assume that the data available provides records of the system functioning at its "natural regime" (e.g., the train network without disruptions) and data on cases where perturbations took place. The inference problem is how information concerning perturbations, with particular covariates such as location and time, can be generalized to predict the effect of novel perturbations. We approach this problem from the point of view of a mapping from the counterfactual distribution of the system behavior without disruptions to the distribution of the disrupted system. A variant on \emph{distribution regression} is developed for this setup.
BibTeX
@misc{colombo2019counterfactual,
  author = {Nicolò Colombo and Ricardo Silva and Soong Moon Kang and Arthur Gretton},
  title = {{Counterfactual Distribution Regression for Structured Inference}},
  year = {2019},
  eprint = {1908.07193},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1908.07193},
}
Figure