Collective Counterfactual Explanations: Balancing Individual Goals and Collective Dynamics

TL;DR

The paper tackles externalities in counterfactual explanations by introducing Collective Counterfactual Explanations (CCE), which couple individual recourse with a mean-field population dynamics model. By penalizing deviations from equilibrium through a distributional divergence, CCE balances personal recourse costs with population-level impact, mitigating competition costs. The framework is formalized via a map $T$ (or a transport plan $\pi$) and relaxed through unbalanced OT, with scalable algorithms and guarantees, and extended to path-guided and ordered-classifier settings. Empirical results across multiple datasets show that CCE achieves a favorable trade-off between modification cost and competitive efficiency, aligning recourse with collective objectives and data-manifold structure. This approach offers a principled, scalable, and causally aware direction for recourse that accounts for societal-level dynamics.

Abstract

Counterfactual explanations provide individuals with cost-optimal recommendations to achieve their desired outcomes. However, when a significant number of individuals seek similar state modifications, this individual-centric approach can inadvertently create competition and introduce unforeseen costs. Additionally, disregarding the underlying data distribution may lead to recommendations that individuals perceive as unusual or impractical. To address these challenges, we propose a novel framework that extends standard counterfactual explanations by incorporating a population dynamics model. This framework penalizes deviations from equilibrium after individuals follow the recommendations, effectively mitigating externalities caused by correlated changes across the population. By balancing individual modification costs with their impact on others, our method ensures more equitable and efficient outcomes. We show how this approach reframes the counterfactual explanation problem from an individual-centric task to a collective optimization problem. Augmenting our theoretical insights, we design and implement scalable algorithms for computing collective counterfactuals, showcasing their effectiveness and advantages over existing recourse methods, particularly in aligning with collective objectives.

Figures (6)

• Figure 1: Comparing CE methods with a non-linear SVM classifier. (Left) wachter2017counterfactual place all recommendations on the decision boundary $L$. (Right) In contrast, Collective CE moves individuals to more populated areas, which are potentially more resource-rich.
• Figure 2: (Left) An example CCE recommendation. (Center) The standard CE method poses a higher risk of revealing the classifier boundary. (Right) Identification strategies are less effective in uncovering the boundary with the more sophisticated design CCE method.
• Figure 3: (Left) The temporal map showing the flow of each point as it moves toward the recourse target. (Right) The back-and-forth method was applied to estimate the CE map. By leveraging displacement interpolation, the optimal flow is depicted across four time steps, showing the transition of the negative region into the positive region using the Moons dataset.
• Figure 4: Comparison of modification and competition costs across 100 experiments with different random seeds. Bar plots show average values, while error bars represent standard deviations. Our method (red bar) achieves lower competition cost but not the lowest modification cost, as it moves points toward higher-density regions. However, when considering the combined metric of modification cost and competition efficiency, it outperforms baselines, achieving the best trade-off.
• Figure 5: (Left) The blue curve represents the percentage increase in modification cost of CCE relative to standard CE as $\lambda_2$ varies from 0.01 to 0.3. The red curve illustrates the competition cost obtained by $\lambda_2 D_{\chi^2}(T_{\#}\mathbb{P}_\textbf{-} \parallel \mathbb{P}_\textbf{+})$. Both curves are supported with confidence intervals. As expected, there is a trade-off between modification and competition cost measures. (Right) The result of CCE on the Adult dataset with $\lambda_2 = 0.1$.

Theorems & Definitions (19)

• Proposition 1: Collective Counterfactual Explanation
• Proposition 2: Existence of a Plan
• Proposition 3
• Proposition 4: Complexity of \ref{['alg:pg-cce']}
• Proposition 5: Actionable CCE Through Linear Constraints
• Definition 1: Push-forward Measure
• Definition 2: Weak Topology
• Definition 3: Set of Couplings
• Definition 4: Linear Constraints
• Definition 5: The Wasserstein Metric