Developing and Evaluating Graph Counterfactual Explanations with

GRETEL

Mario Alfonso Prado-Romero, Bardh Prenkaj and Giovanni Stilo

• Answer the question: “how should an input instance be perturbed to obtain a desired predicted label?"

• Provide recourse to the users via feedback they can act upon to change the prediction result in their favor.

• Help to indetify bias and increase fairness.

• No well-established datasets

• No standard evaluation metrics

• Distinct oracles, built on different frameworks, are used for the same datasets

• Doesn’t compare exhaustively with other state-of-the-art methods

• Framework for evaluating and developing GCE methods

• Open source with modular and extensible design

• Integrates state-of-the-art datasets, oracles, explainers and evaluation metrics