GLANCE: Global Actions in a Nutshell for Counterfactual Explainability

TL;DR

GLANCE tackles global counterfactual explanations by formulating GCEs as a small, interpretable set of actions that maximize population recourse while minimizing average cost and preserving coverage. It introduces a two-phase agglomerative approach that clusters in both feature and action spaces to generate diverse candidate actions and then merges clusters to select $s$ final actions, balancing effectiveness and cost under a size constraint. Theoretical hardness results motivate the need for scalable heuristics, while extensive experiments across five datasets and three model types show GLANCE Pareto-dominates baselines in most cases, with a human user study confirming interpretability benefits of compact action sets. The work demonstrates GLANCE’s robustness, efficiency, and practical impact for fair, explainable deployment of ML systems.

Abstract

The widespread deployment of machine learning systems in critical real-world decision-making applications has highlighted the urgent need for counterfactual explainability methods that operate effectively. Global counterfactual explanations, expressed as actions to offer recourse, aim to provide succinct explanations and insights applicable to large population subgroups. High effectiveness, measured by the fraction of the population that is provided recourse, ensures that the actions benefit as many individuals as possible. Keeping the cost of actions low ensures the proposed recourse actions remain practical and actionable. Limiting the number of actions that provide global counterfactuals is essential to maximizing interpretability. The primary challenge, therefore, is to balance these trade-offs--maximizing effectiveness, minimizing cost, while maintaining a small number of actions. We introduce $\texttt{GLANCE}$, a versatile and adaptive algorithm that employs a novel agglomerative approach, jointly considering both the feature space and the space of counterfactual actions, thereby accounting for the distribution of points in a way that aligns with the model's structure. This design enables the careful balancing of the trade-offs among the three key objectives, with the size objective functioning as a tunable parameter to keep the actions few and easy to interpret. Our extensive experimental evaluation demonstrates that $\texttt{GLANCE}$ consistently shows greater robustness and performance compared to existing methods across various datasets and models.

Figures (5)

• Figure 1: A toy example depicting two negative instances $x_1, x_2$, and five actions. (a) The feature space; the line is the decision boundary. (b) The action space; $l_1, l_2$ depict the decision boundary from the perspective of $x_1, x_2$, respectively.
• Figure 2: Comparison of GCEs on the COMPAS dataset using an XGBoost model. (a) GCEs generated by GLANCE (s=3) and GLOBE-CE (s=3 & default s=67). (b) Effectiveness–recourse cost curves showing the share of individuals (y-axis) achieving recourse below a given cost (x-axis); the area under the curve reflects the average recourse cost. Costs follow Sec. 5.
• Figure 3: Intuition behind clustering approaches. (a) First, GLANCE generates diverse candidate actions from the centroids of feature-based clusters. (b) Then, GLANCE merges clusters based on similarity in both feature and action space, grouping instances that may be further apart but can be explained by similar actions. (c) Clustering solely in feature space can yield suboptimal global actions, trading high effectiveness for high cost or low cost for low effectiveness.
• Figure 4: Comparison of effectiveness ($\mathop{\mathrm{\mathsf{eff}}}\nolimits$) and average recourse cost ($\mathop{\mathrm{\mathsf{avc}}}\nolimits$), normalized with the maximum cost achieved in each dataset/model combination) for the solution of $s$-GCE with $s=4$. Standard deviations are represented by error bars. The red horizontal lines represent the $\mathop{\mathrm{\mathsf{eff}}}\nolimits>80\%$ threshold for evaluating the practicality of the solutions.
• Figure 5: Comparison of effectiveness ($\mathop{\mathrm{\mathsf{eff}}}\nolimits$) and average recourse cost ($\mathop{\mathrm{\mathsf{avc}}}\nolimits$), normalized with the maximum cost achieved in each dataset/model combination) for the solution of $s$-GCE with $s=8$. Standard deviations are represented by error bars. The red horizontal lines represent the $\mathop{\mathrm{\mathsf{eff}}}\nolimits>80\%$ threshold for evaluating the practicality of the solution.

Theorems & Definitions (2)

• Theorem 2: NP-hardness
• proof : Proof of Theorem 2