Revisiting CounteRGAN for Counterfactual Explainability of Graphs

Mario Alfonso Prado-Romero, Bardh Prenkaj, Giovanni Stilo

Keywords: graph counterfactual explainability, generative explainability, graph-to-image

TL;DR: An adaptation of CounteRGAN to consider graph-to-image transformations and generate valid counterfactuals

Abstract: Counterfactual explainability (CE) has been widely explored in various domains ranging from medical image diagnosis to self-driving cars. Graph CE (GCE), on the other hand, and especially, generative-based GCE has yet to be explored. Here, we adapt CounteRGAN, an image-based generative approach, to consider graph adjacency matrices as special black-and-white images and sample valid counterfactuals directly from the learnt latent space probabilistic distribution.