A New Approach to Backtracking Counterfactual Explanations: A Unified Causal Framework for Efficient Model Interpretability
Pouria Fatemi, Ehsan Sharifian, Mohammad Hossein Yassaee
TL;DR
BRACE introduces a unified, causally grounded framework for counterfactual explanations that is computationally efficient and scalable. By optimizing a single objective that balances proximity in the observed space $d_X$ and latent-space changes $d_U$ under a bijectively generated SCM, it can recover multiple established paradigms (e.g., CF, deep backtracking, and causal recourse) as special cases. The approach leverages an invertible mapping $\mathbf{F}$ between latent $\mathbf{U}$ and observed $\mathbf{X}$, enabling backtracking to generalize interventional counterfactuals and to provide realistic, actionable recommendations. Empirical results on synthetic bank-loan data and real German Credit data demonstrate that Brace yields more interpretable and actionable counterfactuals while remaining robust to causal-function perturbations and scalable to larger models.
Abstract
Counterfactual explanations enhance interpretability by identifying alternative inputs that produce different outputs, offering localized insights into model decisions. However, traditional methods often neglect causal relationships, leading to unrealistic examples. While newer approaches integrate causality, they are computationally expensive. To address these challenges, we propose an efficient method called BRACE based on backtracking counterfactuals that incorporates causal reasoning to generate actionable explanations. We first examine the limitations of existing methods and then introduce our novel approach and its features. We also explore the relationship between our method and previous techniques, demonstrating that it generalizes them in specific scenarios. Finally, experiments show that our method provides deeper insights into model outputs.