Counterfactual Metarules for Local and Global Recourse

TL;DR

The paper addresses the need for scalable, interpretable counterfactual explanations that capture both local and global recourse. It introduces T-CREx, a model-agnostic method that learns hyperrectangular counterfactual rules and metarules via tree-based surrogates to summarize recourse options efficiently. It demonstrates competitive or superior performance to baselines across multiple CE desiderata, while offering orders-of-magnitude faster runtimes and applicability to classification and regression. By enabling rapid per-instance explanations through rule lookup and providing global recourse summaries, the approach supports auditing, fairness analysis, and practical decision support in real-world settings.

Abstract

We introduce T-CREx, a novel model-agnostic method for local and global counterfactual explanation (CE), which summarises recourse options for both individuals and groups in the form of human-readable rules. It leverages tree-based surrogate models to learn the counterfactual rules, alongside 'metarules' denoting their regions of optimality, providing both a global analysis of model behaviour and diverse recourse options for users. Experiments indicate that T-CREx achieves superior aggregate performance over existing rule-based baselines on a range of CE desiderata, while being orders of magnitude faster to run.

Figures (12)

• Figure 1: Application of T-CREx to a binary classifier, producing CEs for the red class . Shown are two rules (dark green/orange) for which $\geq 90\%$ of contained points have the alternative blue class . The rules are paired with metarules denoting regions of the input space where each rule is an optimal CE (light green/orange), which can be interpreted as follows. For inputs with $x_1\leq 3$ and $x_2\leq 5$, the orange rule is optimal because it requires changing only one feature (sparsity). Elsewhere, the green rule is preferred because it contains a greater number of points (feasibility). The two rules and three metarules can be combined to create a global textual summary of all recourse options (shown on the right), which in turn enables the near-instantaneous generation of a CE for any single instance (e.g. A or B) via a simple lookup.
• Figure 2: The seven steps of the T-CREx algorithm.
• Figure 3: Performance of T-CREx as a function of the number of trees and $\tau$ (arrows indicate 'better' direction for each desideratum).
• Figure 4: Comparative evaluation of T-CREx0.9, T-CREx0.99, AReS, LORE and RF-OCSE on nine binary classification datasets.
• Figure 5: Distribution of counterfactual distances for all methods.