Coverage-Validity-Aware Algorithmic Recourse
提出一种模型无关的框架，生成对模型更新鲁棒的算法追索方案，利用极小极大概率机理论确保追索在模型变化后仍有效，适用于信用评分、招聘算法等需要公平和透明度的场景。
In “Coverage-Validity-Aware Algorithmic Recourse,” Bui, Nguyen, Yue, and Nguyen introduce a novel framework to enhance the explainability and robustness of algorithmic recourse in machine learning models. Algorithmic recourse provides actionable steps for individuals to achieve favorable outcomes from predictive models, addressing ethical and transparency challenges. However, existing methods often fail to account for the evolving nature of predictive models, rendering current recourses invalid as models update with new data. To address this, the authors propose a novel model-agnostic framework designed to generate recourses robust to future model shifts. The framework constructs a coverage-validity-aware linear surrogate of the nonlinear (black box) model for recourse generation, enabling recourses that remain valid across model updates. The approach leverages theoretical insights from minimax probability machines, demonstrating how varying covariance robustness recovers widely used regularizations, such as ℓ 2 regularization and class reweighting. This method holds promise for applications in credit scoring, hiring algorithms, and other domains where fairness, transparency, and adaptability are critical to maintaining trust and accountability in artificial intelligence systems.