CS&E Colloquium - The Abundance of Great Models: A Blessing or a Curse?

Abstract

The Markov boundary of a response variable T is the minimal set of features that renders all other variables in the dataset statistically independent of T. While some data distributions admit a unique Markov boundary, others contain multiple distinct Markov boundaries for a response variable. The existence of multiple Markov boundaries results in the presence (and in some datasets, the abundance) of multiple predictively equivalent and optimal models. In this talk, I will introduce the foundational theory of multiple Markov boundaries and explore the potential mechanisms underlying their appearance in complex datasets, focusing specifically on biomedical applications. Furthermore, I will discuss the downstream implications of multiple Markov boundaries for predictive modeling, causal discovery, and model translation into real-world decision support systems.

Biography