Generalizability vs. Counterfactual Explainability Trade-Off

2025/05/29 by Veglianti, Fabiano, Giorgi, Flavio, Silvestri, Fabrizio +1
#FOS: Computer and information sciences #Machine Learning (cs.LG)

paper · doi:10.48550/arxiv.2505.23225

Abstract

In this work, we investigate the relationship between model generalization and counterfactual explainability in supervised learning. We introduce the notion of ε-valid counterfactual probability (ε-VCP) -- the probability of finding perturbations of a data point within its ε-neighborhood that result in a label change. We provide a theoretical analysis of ε-VCP in relation to the geometry of the model's decision boundary, showing that ε-VCP tends to increase with model overfitting. Our findings establish a rigorous connection between poor generalization and the ease of counterfactual generation, revealing an inherent trade-off between generalization and counterfactual explainability. Empirical results validate our theory, suggesting ε-VCP as a practical proxy for quantitatively characterizing overfitting.

Citations

• Eliminating Catastrophic Overfitting Via Abnormal Adversarial Examples Regularization
• The Dimpled Manifold Model of Adversarial Examples in Machine Learning
• CF-GNNExplainer: Counterfactual Explanations for Graph Neural Networks
• CounteRGAN: Generating Realistic Counterfactuals with Residual Generative Adversarial Nets
• FOCUS: Flexible Optimizable Counterfactual Explanations for Tree Ensembles
• Understanding Adversarial Robustness: The Trade-off between Minimum and Average Margin
• Model-Agnostic Counterfactual Explanations for Consequential Decisions
• Disentangling Adversarial Robustness and Generalization
• Unrestricted Adversarial Examples
• Learning deep representations by mutual information estimation and maximization
• Local Rule-Based Explanations of Black Box Decision Systems
• Lipschitz-Margin Training: Scalable Certification of Perturbation Invariance for Deep Neural Networks
• Exploring Generalization in Deep Learning
• Formal Guarantees on the Robustness of a Classifier against Adversarial Manipulation
• An empirical analysis of dropout in piecewise linear networks
• Early stopping and non-parametric regression: An optimal data-dependent stopping rule

Related

• Generally-Occurring Model Change for Robust Counterfactual Explanations
• Measurable Counterfactual Local Explanations for Any Classifier
• Rethinking Distance Metrics for Counterfactual Explainability
• Countering Overfitting with Counterfactual Examples
• Finding Regions of Counterfactual Explanations via Robust Optimization