CountARFactuals -- Generating plausible model-agnostic counterfactual explanations with adversarial random forests

TL;DR

This work tackles the challenge of generating plausible, model-agnostic counterfactual explanations for mixed tabular data. It introduces countARFactuals, which leverages adversarial random forests (ARFs) to model data density and efficiently sample realistic counterfactuals. Two algorithms are proposed: (i) integrating ARF with the multi-objective counterfactual explanations (MOC) framework, and (ii) using ARF as a standalone counterfactual generator. Experiments on synthetic data and a real coffee-quality dataset demonstrate improved plausibility and faster generation with only modest trade-offs in proximity and sparsity, highlighting ARFs as a practical tool for recourse in tabular domains.

Abstract

Counterfactual explanations elucidate algorithmic decisions by pointing to scenarios that would have led to an alternative, desired outcome. Giving insight into the model's behavior, they hint users towards possible actions and give grounds for contesting decisions. As a crucial factor in achieving these goals, counterfactuals must be plausible, i.e., describing realistic alternative scenarios within the data manifold. This paper leverages a recently developed generative modeling technique -- adversarial random forests (ARFs) -- to efficiently generate plausible counterfactuals in a model-agnostic way. ARFs can serve as a plausibility measure or directly generate counterfactual explanations. Our ARF-based approach surpasses the limitations of existing methods that aim to generate plausible counterfactual explanations: It is easy to train and computationally highly efficient, handles continuous and categorical data naturally, and allows integrating additional desiderata such as sparsity in a straightforward manner.

Figures (5)

• Figure 1: (a) Proximity and plausibility can be conflicting objectives dandl2020multi; enforcing proximity may lead to unrealistic counterfactuals. (b) To have high proximity (i.e., low $o_{\text{prox}}$ in \ref{['eq:moc-prox']}) and high plausibility (i.e., low $o_{\text{plaus}}$ in \ref{['eq:moc-plaus']}, with $k = 1$), the counterfactual may be in a low-density region.
• Figure 2: Boxplots of the plausibility, proximity ($1 - o_{\text{prox}}$), sparsity ($1 - o_{\text{sparse}}$), hypervolume, number of counterfactuals and runtime for each method and dataset. Higher values are better, except for runtime.
• Figure 3: Exemplary countARFactuals for an instance of bad coffee quality. Arrows indicate changes in comparison to $\mathbf{x}^*$, i.e., a feature's value increase $\uparrow$, decrease $\downarrow$ or change in category $\leftrightarrow$.
• Figure 4: Median empirical attainment function lopez2010 for the negative plausibility and negative proximity. Lower values are better.
• Figure 5: Median empirical attainment function lopez2010 for the negative plausibility and negative sparsity. Lower values are better.