Toward practical and plausible counterfactual explanation through latent adjustment in disentangled space

Abstract

Extensive research into eXplainable AI (XAI) has raised interest in generating counterfactual (CF) explanations. In the past, minimizing the perturbation of input was considered a priority aspect of CF for the benefit of user practicality. However, closeness to the CF data manifold, indicating plausibility, is now emerging as another important property of CF. Thus, we propose a novel framework for generating practical and plausible CFs by minimally perturbing the semantic information of inputs in a disentangled latent space of a generative adversarial network (GAN). Considering the possibility of linear change of semantic information in a disentangled latent space, we obtain the desired CFs using proposed algorithms that adjust the input latents and reference CF latents derived using an optimization-based GAN inversion method. The results of qualitative and quantitative experiments on several datasets from different domains demonstrate the superiority and versatility of our framework. In comparative experiments, it not only achieves 1.0 Validity for test samples from all datasets but also achieves the minimum values of 0.07 Dissimilarity, 5.96 Rec. Error, 0.94 IM1, and 0.01 Infer. Time for the MNIST dataset.