Author: Tien VoPublisher:
ISBN:
Category :
Languages : en
Pages : 0
Book Description
Large-scale hypothesis testing is very important for assessing population differences from sampled data in various application domains. In many cases, high-dimensional data are naturally associated with a graphical architecture, in which measured variables reside on graph vertices and the connectivity of the graph conveys information about the underlying relational structure among the data. Essentially, each edge in the graph represents the relationship between values at its endpoints due to some conceptual dependency, e.g temporal, spatial, functional, anatomical, etc. Available large-scale testing methods often consider dependencies a nuisance, and, by using sufficiently simple, unit-level test statistics, aim to control false discovery rate in a way that is robust to details of such dependence. Where some available methods do incorporate models of dependence, they are limited in scope and they do not take advantage of the graphical nature of the data structure. Given shortcomings of available methods and the importance of the largescale testing problem, we propose a new methodology to incorporate graphical information for hypothesis testing. Our proposed method, graph-based mixture model (GraphMM) is a semiparametric empirical Bayesian approach, motivated from a hybrid procedure that exploits grouping information of model parameters to increase testing sensitivity. We conduct experiments on a parallel computing platform and apply model in the context of a neuroimaging task to detect subtle changes from magnetic resonance imagery.