Variational approaches to the analysis of array data
Yiming Ying, Peng Li, Colin K. Campbell · 2007
m Bayesian mixture model m Heuristic view of variational inference m Experiments and resultsMotivation � The biomedical sciences are generating large datasets Frequently, the interpretation of these datasets involves unsupervised learning. With cancer array data, for example, the use of unsupervised learning can enable resolution of the data into tentative subtypes, which can have differing clinical outcomes � Examples are data from expression arrays, exon arrays, microRNA arrays, etc. Continued I: Motivation � Application of Latent Process Decomposition to a prostate cancer expression array dataset from 100 patients (Glinsky et al, 12625 attributes/probes per sample). Cross validation study on the hold-out log-likelihood indicated 3 principal subtypes, with each subtype having a distinct profile in terms of recurrence or non-recurrence of the disease � Elucidating genes with most significant over- or under-expression within each subtype. −600 −610 −620