[Topic-models] Comparing Topics to Real Classes

Ivan Savov ivan.savov at gmail.com
Mon Aug 1 12:55:06 EDT 2016


Hi Flavio,


 How many topics do you usually run the model with?


A good rule of thumb is to fit a topic model with 2x the number of topics
as the real life classes you want to find. The reason is around 50% of
the topics LDA will learn are "junk topics."

This rule may not apply if you're using a supervised algorithm based on the
\thetas,
so you should try 2x, 3x, the number of classes, and maybe even more.


Ivan
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.cs.princeton.edu/pipermail/topic-models/attachments/20160801/66274e75/attachment.html>


More information about the Topic-models mailing list