Is it possible to set the initial topic assignments for scikit-learn LDA?

Question

Instead of setting the topic_word_prior as a parameter, I would like to initialize the topics according to a pre-defined distribution over words. How would I set this initial topic distribution in sklearn's implementation? If it's not possible, is there a better implementation to consider?

Answer 1

If you have a predefined distribution of words in a pre-trained model you can just pass a bow_corpus through that distribution as a function. Gensims LDA and LDAMallet can both be trained once then you can pass a new data set through for allocation without changing the topics.

Steps:

1. Import your data
2. Clean your data: nix punctuation, numbers, lemmatize, remove stop-words, and stem

3. Create a dictionary

   dictionary = gensim.corpora.Dictionary(processed_docs[:])
   dictionary.filter_extremes(no_below=15, no_above=0.5, keep_n=100000)
   

4. Define a bow corpus

   bow_corpus = [dictionary.doc2bow(doc) for doc in processed_docs]
   

5. Train your model - skip if it's already trained

   ldamallet = gensim.models.wrappers.LdaMallet(mallet_path, 
               corpus=bow_corpus, num_topics=15, id2word=dictionary)
   

6. Import your new data and follow steps 1-4

7. Pass your new data through your model like this:

     ldamallet[bow_corpus_new[:len(bow_corpus_new)]]
   

8. Your new data is allocated now and you can put it in a CSV