Read this

[amritbhanu]

Read this

[amritbhanu]

do people tune lda parameters or even check stability?

• Yes they do that, and in some of the papers they try to have good topic coherence. Not the word "stability". But it is in very initial phase, not much work seen.
• Papers:
  • https://engineering.purdue.edu/RVL/Publications/Rao2011Retrieval.pdf
  • http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.395.7889&rep=rep1&type=pdf
  • http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.380.7196&rep=rep1&type=pdf

[timm]

But it is in very initial phase, not much work seen.

ok. now you've read the papers...

• how to measure "coherence"
• list 10 ways to improve them sorted by ease of implementation, who has tried what before, what might work best

[amritbhanu]

Will have to code the LDA from scratch so that we have flexibility. This will make any implementation easier.
Possible Approaches sorted by ease of implementation are:

• Cohesion (intra) and separation (inter).
• direct appraoch, asking people about topics using Amazon Mechanical Turk #10
• Perplexity shows how well topic-word and word-document distributions predict new test samples.
• Computing the harmonic mean of posterior distribution.
• pointwise mutual information (PMI) between the topic words
• Each term was represented as a vector in a semantic space, with topic coherence calculated as mean pairwise vector similarity. (Word2vec)
• symmetric Kullback–Leibler divergence