Section: New Results
Latent Variable Models for Expert Finding
Latent variable models are widely used in many areas of data analysis: speech recognition, image restoration, finance forecasting. In document analysis one can highlight two methods uncovering hidden structure of the documents. The PLSA method proposed by T. Hoffman  and topic models introduced by D. Blei et al.  acquired popularity and have proved to be very useful statistical description of document corpus. Latent variable models can be directly applied to expert finding task because it implies existence of "hidden" experts who create "observed" documents. Therefore the generative model of a document includes modeling hidden author's contribution to a document based on observed set of authors and modelling hidden expert profiles based on observed words in a document. In previous work  inference in this kind of models was done using EM algorithm. Here we note the empirical nature of this point estimation and in contrast we take a fuller Bayesian approach when prior knowledge is used for parameter estimation. In this case variational Bayesian methods are appropriate for relatively fast and reliable inference.