12
$\begingroup$

There is a new interesting page, called TagOverflow (I came to it from math SE Meta). Its results for Cross Validated are here. Or, copied in here:

enter image description here

For discussion here:

  1. What can this teach us about CV?
  2. What methods were used in TagOverflow? Can we do it better?
  3. Should this (or something like it) be included in SE network?
$\endgroup$
  • $\begingroup$ It seems to be more interesting/adequate with 100 top tags instead of 64. Also, note that the exact geometry of the network is different on each reload (topology is always the same of course, but the 2D embedding seems to be stochastically optimized). $\endgroup$ – amoeba says Reinstate Monica Feb 13 '15 at 15:41
  • 2
    $\begingroup$ Author here. I wanted to post it but you were quicker :). It is a follow-up of Tag Map for CV. Some of description is there: github.com/stared/tagoverflow#tagoverflow. $\endgroup$ – Piotr Migdal Feb 13 '15 at 21:38
  • $\begingroup$ @PiotrMigdal is the warning regarding being blocked real though? $\endgroup$ – Firebug Aug 3 '16 at 13:30
  • $\begingroup$ @Firebug Bad news: yes. Just try doing it a few times in a row and you will get blocked. Good news: but just for a few minutes. $\endgroup$ – Piotr Migdal Aug 5 '16 at 11:24
13
$\begingroup$

This is the output from a social network analysis. (I don't know much about SNA, but we have a tag for it, .) Specifically, this is an undirected graph where the sizes of the nodes are proportional to the number of questions with that tag (actually, it's probably the log or square root, etc., of the number [no it isn't, note the comment from @PiotrMigdal below]), and the weights of the edges is proportional to the number of questions with both tags. By default, the colors are the result of community detection, a kind of clustering of nodes in networks. I notice that you can change this, however. Importantly, notice that by default these are only the top 64 tags (although this can be changed as well).

Here are some specific answers:

  1. I think this is pretty neat. It gives us one way of exploring the structure of the threads on CV. Under the (I think reasonable) assumptions that most of the threads are tagged appropriately and that the tags represent meaningful topics within statistics and machine learning, this tells us about the volume of interest (i.e., number of questions) in different topics within our purview and their interrelationships. There seem to be five primary clusters, for example.

    a. A $\color{blue}{\rm modeling}$ cluster, anchored by and
    b. A $\color{purple}{\rm basic\ statistics}$ cluster, anchored by and
    c. A $\color{green}{\rm machine\ learning}$ cluster, anchored by and
    d. A $\color{red}{\rm somewhat\ more\ mathematical\ take\ on\ statistics}$ cluster, anchored by and
    e. A $\color{orange}{\rm time\ series}$ cluster, composed entirely (at least within the first 64 tags) of and

  2. The analytical methods used were based in social network analysis, no doubt built on a great deal of SQL [no it isn't, note the comment from @PiotrMigdal below]. Maybe someone here could do it better, but this was pretty sophisticated and I notice that there are only two CV users who are listed as having answered more than one question on the [S-N] tag.

  3. SE could adopt this as part of, or something similar to, its data site, if the developers wanted.

$\endgroup$
  • 5
    $\begingroup$ One of the authors has participated on CV, Piotr Migdal and on a few other stackexchange sites. One of the projects I think would be neat would have questions as nodes, and then see what highly linked questions could potentially be FAQ's. $\endgroup$ – Andy W Feb 13 '15 at 19:42
  • $\begingroup$ @AndyW Nice ideas. I don't have that much time, but if there is a need, we can make a statistical analysis of this site (I launched a collaborative pad: hackpad.com/metastatsis-4gfEVdu8z7g). $\endgroup$ – Piotr Migdal Feb 14 '15 at 14:14
  • 2
    $\begingroup$ @gung area ~ counts (no log); queries from JSON API, not SQL (at least, not directly); see description for more information. Here community detection is mostly for visual purposes, I can get it more serious (with a bit different graph and method) - of course, the most resilient things will stay the same. $\endgroup$ – Piotr Migdal Feb 14 '15 at 14:17
  • 1
    $\begingroup$ @AndyW: As an aside, "highly linked questions" are directly available via the "frequent" tab. $\endgroup$ – amoeba says Reinstate Monica Feb 14 '15 at 14:55
  • 1
    $\begingroup$ @PiotrMigdal, feel free to add your own answer as the authoritative post. You needn't refrain just because I already posted. I could even delete mine, if necessary. $\endgroup$ – gung - Reinstate Monica Feb 14 '15 at 15:41
  • $\begingroup$ @gung Please don't delete it. I really like seeing how things are getting interpreted. In any case, I don't have much to add except for things I have already said/ $\endgroup$ – Piotr Migdal Feb 14 '15 at 16:44

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .