Author Tree-Structured Hierarchical Dirichlet Process.

Three key aspects of online discussion venues are the multitude of participants, the underlying trends of content, and the structure of the venue. However, most models are unable to take into account all three of these. In hierarchically organized message forums, authors may participate differently at multiple levels of sections, with different interests and contributions across the hierarchy. Well-designed probabilistic models of online discussion are applicable to many tasks such as prediction of future content or authorship attribution. However, traditional models such as Hierarchical Dirichlet Processes (HDPs) do not fully take into account authors, and are further unable to fully take into account deep hierarchical venues where documents can arise at all tree nodes. We introduce the Author Tree-structured Hierarchical Dirichlet Process (ATHDP), allowing Dirichlet process based topic modeling of both text content and authors over a given tree structure of arbitrary size and height. Experiments on six hierarchical discussion data sets demonstrate better performance of ATHDP compared to traditional HDP based alternatives in terms of perplexity and authorship attribution accuracy.