What does it mean for Latent Dirichlet Allocation results to be "good"? In most paper, Latent Dirichlet Allocation (LDA) model is used for clustering, and the value of $K$ is trained manually (e.g. http://astro.temple.edu/~tua95067/grbovic_cikm.pdf). They claim that this value has a good balance between total and individual coverage.


*

*I don't know how to define the "total coverage" and "individual coverage".

*Is there a good method to train the value of K, not manually?

 A: K is the number of topics for your model. Just like in any other clustering problem, choosing the right number of clusters is very problem/data-specific and usually is chosen manually/randomly.
To estimate the number of clusters K, non-parametric methods can be used, for example, the Hierarchical Dirichlet Process (HDP) described by Wang et al. 
In the paper you link to, however, (at a glance) it seems like they are trying to find a number of topics K that evenly distributes their instances ("e-mails") across all folders. Note that this last point about folders is not LDA-specific!


*

*their coverage for a topic k is the number of emails for that topic associated to that folder;

*next they require that a topic should cover at least 50% of all emails in that folder to be dominant (in that folder)


Say you have 10 emails in 2 folders X and Y and do an LDA with a K of three (topics A-C) and you get:


*

*6 mails in folder X: 3 A, 2 B, 1 C

*4 mails folder Y: 1 A, 1 C, 2 C


NB: I could not confirm that each email is only assigned to one folder, but I assume this is the case.
Now they define total coverage as the traffic (i.e., the total proportion of emails) where their topic covers 50% of all mails in a folder. In our case we get:
For topic A, the individual coverage is 30%:


*

*in folder X, it is dominant/has 50%, and represents 30% of all mail

*in folder Y, it has less than 50%, so we ignore it


For topic B, the individual coverage is 0%:


*

*is not dominant in any folder


For topic C, the individual coverage is 20%:


*

*50% in folder Y, accounting for 20% of all mail


Thus, for this outcome, total coverage would be 30% plus 20% = 50%.
Their strategy now is to find an "optimal" K that maximizes this total coverage while also ensuring high individual (per-topic) coverages.
This means, their method of defining K uses "external information" not part of the LDA process itself, namely the folder-assignments of emails by users (which are independent of the topic assignments by the LDA).
