Big Data Computing

521Big Textual Data Analytics and KnowledgeManagement

document generation. We need extra notation here for the universe of all

possible topic-term distributions. Recall that this is actually the universe of

all possible topics. For a nite set of m topics, this universe corresponds to a

simplex in m + 1 space. As the HDP approach can account for innitely many

topics, let us denote the corresponding universe by T.

It is also easier to understand HDP, if we rst look at the implied topic mod-

eling within a given document and then generalize to a hierarchical model.

So, rst, on the level of one document d

j

, a DP can be used as follows:

1. On the corpus level, select a prior probability distribution G

0

on the

set of topic-term distributions ...