A0594
Title: Learning from heterogeneous data with stick-breaking variational autoencoder
Authors: Jaeyoung Lee - Virginia Tech (United States) [presenting]
Hongxiao Zhu - Virginia Tech (United States)
Abstract: Modern data often exhibit intricate heterogeneous structures caused by diverse data sources, subpopulations, nested experimental designs, or other unknown factors. Conventional statistical models are inadequate for such data as they frequently overlook the inherent heterogeneous structure. An innovative statistical learning framework is introduced to capture latent heterogeneous structures among samples while facilitating prediction and association analysis. Specifically, the framework employs a stick-breaking variational autoencoder to characterize the heterogeneous data structure and link the latent stick-breaking process with a response variable. The advantages of modeling latent heterogeneous structures are illustrated through simulations and a real data application involving brain tumor images.
