Thats not true in many cases. The wholesale price cap is 7.70€ /GB. This is the price telcos have to pay other telcos for their users roaming. And this is more expensive than end-user price in the nordics/baltic region.
One way is to learn clusters in the data over the whole time period and then calculate the cluster distribution of the same clusters for each time interval separately. You can then track the proportion of each cluster over time as a time series.
So for example if you have a word cluster that describes computers, you could see it start growing in the seventies, while having near-zero proportion in 19th century etc.
Not sure what you mean by a new classification approach. There is no classification here, since there are no labeled documents. This is purely unsupervised topic modelling. The topics are mathematical objects. How they are later named or grouped for better human readability is a subjective matter.
I think you're mixing up what topics are. The actual topics as generated by LDA are the concatenated word lists (actually distributions of all words in the corpus, of which i concatenate the top 8 words to generate a meaningful descriptor of the topic). So server-client-http-request-service-ruby-connection-user is one topic / word distribution, in which "ruby" happens to be 6th most probable word, likely because it appears a lot in posts on servers, web services etc. It does not mean ruby the word itself is classified to be server related. Same applies to the other examples you gave.
The categories/domains I simply assigned manually, to show how one could possibly interpret these word distributions that LDA generated.
The correlation between individual upvotes and comments isn't really what's the post is about, it's purely an illustration and has no impact on the topic extraction or interpretation.
For what it's worth, I did check the correlation coefficient between the two sets (it's 0.81)
The topic "language-type-program-code" is the 6th ranked topic out of 30 in terms of comments, so it's pretty high. Considering the error bars, it could possibly be even further up.
There are a few things i feel i need to clear up here.
1) I'm not sure what you mean by being unable to handle nonmonotonic/multivariate effects. There is no issue with nonmonotonic effects, the sum of feature contributions is always how each tree actually predicts. Yes, interpretation would be somewhat harder, but can be solved by looking at feature value and/or distribution once you know its contribution.
2) Mean impurity decrease or Breimans feature permutation based method have almost no use in a setting i'm describing. They are both static measures in the sense they only apply to the model itself, and they will tell you nothing about a particular prediction on a data point or a set of predictions on the data set.
3) The issue with highly corelated features is indeed still there, but it is literally exactly the same problem that mean decrease impurity and Breimans method would face.