The article presents the Structured Counterfactual Diffuser (SCD), a new framework designed to generate counterfactual explanations for black-box neural network models. Counterfactual explanations are a powerful tool for understanding and building trust in these models. The SCD uses diffusion models to learn the underlying data distribution, which it then uses to generate plausible, diverse counterfactuals for any given model, input, and desired prediction. The authors claim that their approach produces counterfactuals that are not only more plausible than those produced by current state-of-the-art methods, but also show significantly better proximity and diversity.
Publication date: 22 Dec 2023
Project Page: Not Provided
Paper: https://arxiv.org/pdf/2312.13616
