Figure 7: Maximum likelihood from incomplete data via the EM algorithm. Here, “theta” denotes θj and “theta correction” denotes the offset j . The 10 topics are obtained by joining the top 5 topics ranked by θjk and another top 5 topics ranked by |jk|, k = 1, · · · , K. Under CTR, an article of wide interest is likely to exhibit more topics than its text exhibits. For example, this article brings in several other topics, including one on “Bayesian statistics” (topic 10). Note that the EM article is mainly about parameter estimation (topic 1), though is frequently referenced by Bayesian statisticians (and scholars in other fields as well).