V-DS logo Visual Data Science Lab FGV EMAp · Rio de Janeiro

Explainable & Fair AI

Methods and visual tools to explain, audit, and de-bias machine learning models

Explainable & Fair AI develops methods and visual tools to understand, audit, and correct the behavior of machine learning models. On the explainability side we work on counterfactual explanations (Pareto-optimal counterfactual antecedents, what-if scenarios for crime and student dropout, counterfactual analysis of urban perception), explanations for image classifiers (style augmentation as adversarial attacker and defender), graph neural networks (surrogate-based explanations and topology-aware analysis of explanations in dynamic graphs), and text classifiers (incorporating human rationales to make explanations plausible without losing faithfulness). On the fairness side we study how to identify and mitigate discrimination in data and models, from validating the equity of national exams to enforcing fairness with ensembles of diverse Pareto-optimal models. Many of these methods are applied inside our visualization systems for crime, education, and legal analysis.

Explainable & Fair AI

Researchers

Publications