Reproduce “ACE: Adapting sampling for Counterfactual Explanations” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Margarita A. Guerrero, Cristian R. Rojas
The ACE algorithm generates high-quality counterfactual explanations for black-box classifiers by employing Bayesian optimization and stochastic processes to significantly reduce the required number of model queries. It consistently achieves superior sample efficiency and produces plausible, valid explanations across diverse datasets.
Key insights
- Modeling the black-box classifier's latent function with a Gaussian Process allows for efficient exploration by quantifying prediction uncertainty, reducing the number of costly model queries.
- An adaptive penalty method within a Lagrangian framework effectively pushes candidate counterfactuals towards the decision boundary while satisfying structural constraints like plausibility and sparsity.
- A hybrid optimization strategy, combining continuous optimization for numerical features and Branch and Bound for categorical features, enables robust and sample-efficient CFE generation across diverse data types.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.