Initialization of Dynamic topic model and relation to Static LDA model

[slim oueslati]

Hello,

 I'm using  ldaseqmodel and I have a problem with the prior initialization of the beta distribution (word_topic matrix).  As described in the documentation we can control the initialization of the DTM model throw the 'initialize' parameter. 

Here is my code :

LDA_seq=  LdaSeqModel(corpus, time_slice, id2word, num_topics=20, initialize='own', sstats=LDA_prior)

LDA_prior is the beta distribution of a pre-trained LDA model : 

LDA_prior = np.transpose(LDA.get_topics()) 

with : 

LDA = LdaModel(corpus=pre_corpus, num_topics=10, pre_id2word)

While training the dynamic topic model, I get this logging Info :

In the third line, it is written : ''using symmetric eta at 0.05'' : I didn't expect this output  since I already initialized the first time slice with sstats input parameter !

Could you please explain what is the difference between the two initialization (eta and sstats)?

Another question please :

In the dynamic model, does we train separately on each time slice a static LDA model 

with the corpus contained on that time slice and then we do an update according to the previous state.

Thank you so much !

Slim

[Charlotte Knickrehm]

Hey,

just wondering: did you solve the problem? I am running into the same problem and can't find an answer.

Thank you!