AIM-CF: Fast and Precise Counterfactual Explanations via Approximate Inverse Models

Abstract

Counterfactual explanations (CFs) have become a key technique for validating the decisions of machine-learning models at the level of individual instances. However, conventional search-based algorithms incur (i) high computational cost owing to random or evolutionary exploration and (ii) inflated distances when many one-hot categorical features are present in tabular data. We introduce AIM-CF, a fast and precise CF generator that extends the Approximate Inverse Model Explanation (AIME) framework. AIM-CF consists of three deterministic stages: (1) minimal categorical flipping guided by the global feature importance derived from a representative prototype, (2) axis-wise bisection that shrinks each continuous variable to the decision boundary, and (3) micro-random refinement, which removes residual diagonal slack. The entire procedure requires only forward evaluations of the black-box model and no gradients. On the UCI Heart-Disease dataset, AIM-CF reduces the mean L2 distance from 0.92 (DiCE-Random) to 0.83 while cutting average generation time from 1.78 s to 0.10 s on a CPU—a 17-fold acceleration—and achieves a 100 % success rate versus 97 % for DiCE. These results demonstrate that AIM-CF surpasses the existing methods in terms of distance, speed, and robustness for tabular data containing both categorical and continuous attributes. We conclude by discussing how AIM-CF can be combined with other inverse-problem XAI techniques to provide actionable and explainable decision support in medical and financial applications.

Ähnliche Arbeiten