Convergence test and input count matrix about fasttopics HOT 3 CLOSED

@satriaps Thanks for your interest in fastTopics. To answer your questions:

1. The use of the multinomial measurement model (upon which the topic model is based) is supported by several recent publications, among them Townes et al (2019) and Sarkar & Stephens (2020). It is true some accounting of overdispersion could help in some settings. Most importantly, you should use the "raw" counts, not the log-transformed counts.

2. Yes, that is a feature that I would like to implement. Thank you for the suggestion. One thing I should say though is that people do not typically wait for the optimization to "converge" to a fixed point (however convergence is measured), particularly for large or complex data sets. So my general advice is that monitoring convergence of the optimization is helpful, but, depending on the setting, a good solution may not require convergence.

Hope that helps.

from fasttopics.

@pcarbo (2) is now addressed by c71461e so this issue can be closed.

from fasttopics.

Thanks!

from fasttopics.