PUB - Publikationen an der Universität Bielefeld

We want to identify independent projections of interest inmultivariate data in an unsupervised manner. Dealing withdata of a specific temporal or spatial structure is well es-tablished for this so-called blind source separation prob-lem. The recently published GraDe-algorithm addressesmore complex network structures often found in biology;it analytically separates sources with respect to a givennetwork. We translate these separation assumptions andpropose a flexible Bayesian model to include for instancemissing observations and parameter priors. Technically,we define a Gaussian graphical model with latent variablesand estimate the parameters using expectation maximiza-tion. As large-scale application we consider gene expres-sion data, where the dependence structure is given by aliterature-derived transcriptional regulation network. Wedemonstrate that the model identifies relevant biologicalprocesses.