Graph Explainability and Generation

Graph-learning models are most useful scientifically when we can examine the evidence behind their decisions and determine whether generated graphs preserve the structures that distinguish real network domains.

This ongoing line of work combines explainability and interpretability for graph-based models with reliable evaluation of graph generative models. Current problems include identifying discriminative sub-networks in functional connectomes and evaluating graph generators beyond aggregate summary statistics. This focused page is part of the broader Learning on and with Graphs research thread.