Jun Xie, Ph.D., Professor of Statistics, Purdue University

"A Statistical Model for Causal Feature Representation"

In traditional statistical analysis, including experimental design, features are usually well-defined by a set of random variables. Statistical advancements over the past two to three decades have primarily focused on variable selection techniques. However, modern data analysis frequently deals with a large number of random variables. Features are not clearly defined in the data and need to be learned before we conduct any analysis. Feature representation learning is a well-studied topic in machine learning. My research group focuses on causal feature learning, which is to acquire features for causal inference. In this talk, we demonstrate a statistical methodology known as Sliced Inverse Regression (SIR) and leverage its potential to automatically learn features in two scenarios of causal inference. The first application involves individualized treatment rules, wherein SIR defines a low-dimensional feature space for estimating conditional average treatment effects. In our second application, we use SIR to learn invariant features across multiple data environments, facilitating out-of-distribution generalization.

Stay connected TwitterFacebook LinkedIn YouTubeInstagram