Interpretability and explainability in machine learning

Machine Learning is becoming ubiquitous and data-driven models are increasingly used to make high-stakes decisions in sensitive domains such as healthcare, safety systems, education, and criminal justice. Accordingly, it is important to ensure that decision-makers properly understand how these models work such that they can trust their outputs. The course will cover the following topics. Fundamentals: an overview of the field and seminal papers, the definition of interpretability and explainability, scope of interpretability, properties of explanations, evaluating interpretability. Interpretable Models: interpretability properties of linear regression, logistic regression, decision trees, rule-based techniques, generalized additive models, and instance-based approaches. Model-Agnostic Methods: partial dependence plot, individual conditional expectation, accumulated local effects plot, feature interaction, feature importance, Shapley additive explanations, local and global surrogates. Example-Based Explanations: counterfactual explanations, adversarial examples, prototypes, and influential instances Advanced Topics: Interpretability of Neural Networks and Deep Learning, interpretability and causality, fairness, debugging, human-in-the-loop approaches.