Abstract:
|
Large-scale hypothesis testing is very important for assessing population differences from sampled data in many application domains. Available methods often avoid any reliance on dependence information when constructing the pairwise test statistic. In many cases, data dependencies (e.g. temporal, causal, spatial, etc.) may be informed by available graph structures, in which vertices of the graph are associated with measured variables and edges convey information about potential correlation. With that observation, we investigate a new methodology which is able to accommodate graphical information when performing hypothesis testing. Our approach is motivated from a hybrid approach that combines clustering information to increase testing sensitivity. The proposed graph-based mixture model (Graph-MM) is fitted using a novel auxiliary variable Markov-chain Monte Carlo scheme that takes advantage of spanning trees within the input graph. We investigate computations deployed on a parallel computing platform, in the context of a neuroimaging task to detect subtle changes from magnetic resonance imagery.
|