A0970
Title: Adaptive projected two-sample comparisons for high-dimensional data
Authors: Tianyu Zhang - University of California, Santa Barbara (United States) [presenting]
Abstract: The purpose is to discuss multiple two-sample comparison procedures based on projective scores, which reduce high-dimensional distributions to data-driven, low-dimensional directions. Motivated by structured signals in single-cell genomics data, the methods localize the difference between the two groups to a single direction or a subset of dimensions, enhancing both power and interpretability. One approach incorporates a debiasing step inspired by semiparametric double machine learning, allowing us to construct valid tests for scientifically meaningful projected null hypotheses. Settings, where debiasing is either impractical or unnecessary, are also investigated, developing flexible projection techniques that yield approximately Gaussian statistics even when influence functions are unknown.
