Guidance on topic model selection

[AmosFong1]

Similar to #114.

Loglikelihood does not seem very informative as it is not optimized even at iteration 500. In the vignette, loglikelihood and minmo (topic-coherence) are the recommended metrics for model selection. What are your thoughts on this result, and is loglikelihood still informative in this case?

[SeppeDeWinter]

Hi @AmosFong1

Yes, the selection is not always 100% objective.
What can help is to plot each metric individually. In this plot all metrics are scaled between 0 and 1.
However, the absolute difference between each iteration for some metrics might not be as large as the
relative difference you are plotting here.

I hope this helps?

All the best,

Seppe