kejunxiao / sparse_btm

a cpp implementation of sparse biterm topic model, 10x faster than origin implementation because using sparse-sampler.

sparse_btm

a cpp implementation of sparse biterm topic model, 10x faster than origin implementation because using sparse-gibbs-sampler.

features:

• being suitable to model for user-click-sequenece(Rcommandation System) or short-text(NLP), because it assume that adjacent N-items belong to a topic;
• using sparse-gibbs-sampler, 10x faster than origin implementation;

arguments:

---

Biterm Topic Model (Sparse-Sampler)

---

Parameters:

• -input
  path of docs file, lines of file look like "word1 word2 word3 ... \n"
• -output
  dir of model(topic_biterm_sum, topic_word) file
• -num_topics
  number of topics
• -alpha
  symmetric doc-topic prior probability, default is 0.05
• -beta
  symmetric topic-word prior probability, default is 0.01
• -window_size
  window size for biterms, default is 2
• -num_iters
  number of iteration, default is 20
• -save_step
  save model every save_step iteration, default is -1 (no save)

usage:

./sparse_btm -input short_text.txt -output model_out/ -num_topics 100 -window_size 3 -num-iters 20 -save_step 10

sparse_btm's Issues

Difference from BTM

I’m really excited for this repo because it has the potential to process TBs of text data.

However, I’m interested in the difference between sparse-btm and the original btm.

I’m not sure what the sparse sampler means? And what’s the PMI or coherence difference between the two models?

Any disadvantages of sparse btm?

Many thanks