Reproduce “Interpretable Counterfactual Explanations Guided by Prototypes” on OpenResearch

About this paper

Arnaud Van Looveren, Janis Klaise

A method for generating interpretable and fast counterfactual explanations for classifier predictions, particularly for black-box models, is introduced. It leverages class prototypes to guide the search, ensuring explanations are realistic and class-specific, achieving up to a 100x speed-up for black-box models and improving quantitative interpretability metrics.

Key insights

• Guiding counterfactual explanations with class-specific prototypes significantly enhances their interpretability and realism by ensuring generated instances resemble typical examples of the target class.
• Leveraging prototypes enables a drastic reduction in computational cost for black-box models by obviating the need for prediction-based gradient calculations, leading to substantial speed improvements.
• Meaningful perturbations of categorical features can be achieved by first inferring inter-category distances and then embedding them into a Euclidean space, expanding the applicability of counterfactuals to diverse real-world datasets.

Summary by alphaXiv.