Inference for Interpretable Machine Learning: Feature Importance and Beyond

This lecture from the Colloque des sciences mathématiques du Québec (CSMQ) features Genevera Allen from Columbia University discussing the critical challenge of ensuring trustworthiness in machine learning interpretations. Explore how feature importance and interpretability methods can be verified and trusted when making crucial societal, scientific, and business decisions. Learn about Allen's empirical stability study revealing that feature interpretations are generally less reliable than predictions, and discover a new statistical inference framework for quantifying uncertainty in feature importance and higher-order feature patterns. The presentation introduces a distribution-free approach to test whether features significantly contribute to any machine learning model's predictive ability, demonstrated through scientific case studies and illustrative examples. Particularly valuable for researchers and practitioners concerned with trust, transparency, and accountability in machine learning systems.