Causality for Robust ML

This lecture by the instructor from Oregon State University discusses the indispensability of causality for robust and reliable machine learning. It covers topics such as ideal datasets, crucial problems with missing data, graphical models for encoding assumptions transparently, and the recoverability of missing data. The lecture also delves into problematic structures, testable implications, linear models for interference, and main results related to bias amplification and sample size impact. Theoretical impossibility theorems for missing data are explored, along with real-world examples of missing not at random scenarios. The lecture concludes with references to related works.