h2coder/plda

Is the topic distribution computation in LDASampler::GenerateTopicDistributionForWord correct?

Closed this issue · 0 comments

It looks like the distribution computation in 
LDASampler::GenerateTopicDistributionForWord has:

    distribution->push_back(
        (topic_word_factor + beta_) *
        (document_topic_factor + alpha_) /
        (global_topic_factor + num_topics * beta_));

But based on Algorithm 3.1 in 
http://www.datalab.uci.edu/papers/fastLDA_2008.pdf, shouldn't the 
denominator in the value above be (global_topic_factor + num_words * 
beta_)?

Original issue reported on code.google.com by ritz.kar@gmail.com on 14 Mar 2009 at 7:37