PAR: Plausibility-aware Amortized Recourse Generation

TL;DR

Plausibility-aware Plausible Recourse (PAR) reframes algorithmic recourse as a constrained MAP problem over a tractable data distribution modeled by Probabilistic Circuits (PCs). By training a neighborhood-conditioned recourse generator, PAR delivers high-likelihood counterfactuals that respect feasibility constraints such as immutability, monotonicity, and causality, while maintaining close proximity and sparsity where possible. The method combines explicit plausibility signals from the accepted-class density with classifier feedback, and augments with a local search refinement for robustness and sparsity. Empirical results on Adult, German Credit, and GMSC show PAR achieves strong validity and plausibility with fast generation times, outperforming several baselines and remaining robust to classifier changes. The approach offers scalable, model-agnostic recourse generation that can be reused across users when immutable features are fixed, though it requires explicit constraint specification and incurs inference-time overhead due to local search.</br>All math is denoted with $...$ to reflect the formal MAP objective and likelihood terms used across the method, e.g., $x^+ = \arg\max_{x} \log p^+(x)$ subject to $f(x)=1$ and $x \in \mathcal{F}(x^-)$.

Abstract

Algorithmic recourse aims to recommend actionable changes to a factual's attributes that flip an unfavorable model decision while remaining realistic and feasible. We formulate recourse as a Constrained Maximum A-Posteriori (MAP) inference problem under the accepted-class data distribution seeking counterfactuals with high likelihood while respecting other recourse constraints. We present PAR, an amortized approximate inference procedure that generates highly likely recourses efficiently. Recourse likelihood is estimated directly using tractable probabilistic models that admit exact likelihood evaluation and efficient gradient propagation that is useful during training. The recourse generator is trained with the objective of maximizing the likelihood under the accepted-class distribution while minimizing the likelihood under the denied-class distribution and other losses that encode recourse constraints. Furthermore, PAR includes a neighborhood-based conditioning mechanism to promote recourse generation that is customized to a factual. We validate PAR on widely used algorithmic recourse datasets and demonstrate its efficiency in generating recourses that are valid, similar to the factual, sparse, and highly plausible, yielding superior performance over existing state-of-the-art approaches.

Figures (3)

• Figure 1: Overview of PAR shows the PAR architecture, where class-conditional probabilistic circuits guide a neural recourse generator to produce valid and plausible counterfactuals under structural constraints.
• Figure 2: Effect of local search (LS) on plausibility, sparsity, and similarity across datasets. Each bar pair compares counterfactuals before and after applying LS.
• Figure 3: Robustness to classifier change measured by the mean predicted score $\hat{y}$ of generated counterfactuals under an alternative classifier (mean $\pm$ std across folds). Higher $\hat{y}$ indicates more stable decision flips under model replacement.