29
$\begingroup$

Please post your tag synonym suggestions as new answers in this thread, one answer per suggestion. Upvote answers where you believe that the suggested tags should be made synonyms, and downvote answers where you believe the tags should remain separate. Well upvoted suggestions will be eventually implemented by the moderators (and then the corresponding answers will be deleted).

| |
$\endgroup$
  • 8
    $\begingroup$ @whuber, if this is inappropriate or unhelpful, then I can delete it, but my goal is the opposite: I don't want you (or the other mods) to have to spend time thinking about this & doing a lot of extra research. I'm hoping to bring this to the attention of the community, which, by voting (commenting, etc), will have done all of that for you guys. Nb, w/ respect to lme & multilevel, I'm not sure anyone will have the requisite upvotes--there have only been 3 questions w/ multilevel, according to it's page. $\endgroup$ – gung - Reinstate Monica Jun 7 '12 at 16:59
  • $\begingroup$ Comments are not for extended discussion; this conversation has been moved to chat. $\endgroup$ – whuber Jan 21 '16 at 14:55
4
$\begingroup$

$\leftarrow$

As a piece of software, xgboost is just one (popular) library. However, it's not the only boosting software; we do not have tags for other individual boosting libraries like light-gbm, for instance.

stats.SE is not primarily oriented to debugging issues with statistical software, so it can be confusing to have tags for individual libraries as opposed to the underlying statistical procedures that those libraries implement.

The specific implementation choices that xgboost makes naturally have statistical and computational consequences, and it can be reasonable to ask about them, but I'm not sure we need to dedicate a tag for this boosting implementation (or, in general, every boosting implementation).

| |
$\endgroup$
  • $\begingroup$ It's sometimes hard to draw a clear demarcation between "libraries" (or packages or whatever), statistical software per se, and methods. How about scikit-learn or vowpal-wabbit, for instance? $\endgroup$ – chl Nov 9 at 18:57
  • 1
    $\begingroup$ I appreciate the challenge -- we want tags to put like things together. And, at least in the case of sklearn, it's so widespread as a standard part of the python-language data science software stack that it's almost unavoidable to want to create a tag for those posts. On the other hand, I have similarly mixed feelings about all software tags, including sklearn, because it can create the impression that CV.SE is a support website for software. "Why do you have the tag for __ if my question about a stack trace isn't on topic?" is a common enough complaint when we close a software question. $\endgroup$ – Sycorax Nov 9 at 19:43
  • $\begingroup$ I can understand. sklearn is probably to Python what ggplot2 (or the tidyverse nowadays) is to R ;-) Anyway, these were just some extreme examples, because I also asked myself the question for years. In this particular case (xgboost → boosting), however, I agree with you. $\endgroup$ – chl Nov 9 at 20:07
  • $\begingroup$ I'll make this change before the New Year if no one objects. $\endgroup$ – Sycorax Nov 27 at 2:06
3
$\begingroup$

$\leftarrow$

We have a tag for [segmentation] that seems synonymous to clustering. Consider their excerpts:

In marketing/economics, the task of dividing some population such as customers into sub-groups based on some type of shared characteristics or demands. Statistically, various unsupervised and supervised methods may be used for that (clustering, conjoint analysis, tree, etc.).

Cluster analysis is the task of partitioning data into subsets of objects according to their mutual "similarity," without using preexisting knowledge such as class labels. [Clustered-standard-errors and/or cluster-samples should be tagged as such; do NOT use the "clustering" tag for them.]

The accepted (and highest voted) answer to the question Differences between clustering and segmentation states they are "two faces of the same coin".

Many of the existing threads categorized under [segmentation] seem like they would do fine under [clustering], although some threads appear to use "segmentation" in a completely different way than clustering or segmentation's excerpt. I think the different uses should be un-/re-tagged, and [s] made a synonym of [c]. Here are some stats:

  • [clustering]: 3555 threads, excerpt, full wiki
  • [segmentation]: 129 threads (of which 34 have both tags), excerpt, no wiki
| |
$\endgroup$
  • 2
    $\begingroup$ (+1) Agree. Segmentation is often used in place of clustering in market, business or finance research (which are other domains than the one mentioned in the accepted answer you referred to). Maybe a clarification in the wiki after the merge would be helpful. $\endgroup$ – chl Nov 20 at 11:15
  • $\begingroup$ If there are no protests, I start doing this after 29-11-2020 $\endgroup$ – kjetil b halvorsen Nov 22 at 18:51
  • $\begingroup$ @gung: Looking more at this, the seg posts seems to be two clusters: clustering, and more like image-segmentation (or the same for signals/time series.) I'm unsure if it is correct to classify that as clustering? $\endgroup$ – kjetil b halvorsen Nov 26 at 16:01
  • $\begingroup$ @kjetilbhalvorsen, image segmentation could be folded under the umbrella of clustering, but it isn't what we prototypically mean by clustering. Since I gather we have a tag for that, it would be best to retag those threads as [i-s], & then make the mapping (& merge & delete...). $\endgroup$ – gung - Reinstate Monica Nov 26 at 17:45
  • $\begingroup$ @gung: Is it OK then to include signals&time series (1D images) under that tag? $\endgroup$ – kjetil b halvorsen Nov 26 at 18:09
  • 2
    $\begingroup$ @kjetilbhalvorsen, I don't really know. We're at the edge of my expertise. You might ask Sycorax, since he seems more knowledgeable. I suppose what I would say is that prototypically, clustering is grouping patterns / records into sets; segmentation is partitioning adjacent elements of a single image into subregions that belong to the same thing. You can cluster multiple time series into sets that are similar, or you can dice up a single time series, say a recording, into person A speaking / person B speaking. $\endgroup$ – gung - Reinstate Monica Nov 26 at 19:10
  • 1
    $\begingroup$ I think that "segmentation" is a word that's used in different ways for different tasks/specializations. "Image segmentation" is a term of art about how to process images (it can be either supervised, such as using a mask, or unsupervised, which detects edges and textures to allocate pixels) and is distinct from "(customer) segmentation" in marketing. AFAIK, "(customer) segmentation" is unsupervised or exploratory, and otherwise known as "clustering" or "EDA" but this is just my impression as a person who's heard about it in passing. $\endgroup$ – Sycorax Nov 27 at 2:02
  • 1
    $\begingroup$ I don't think that we should include signal/time series under [image segmentation], because to me, it's surprising to call that data an "image," in my humble opinion. I don't know how much time series segmentation (dividing up a single time series) or time series clustering (grouping together two or more time series) shares in common with the practice of image segmentation. $\endgroup$ – Sycorax Nov 27 at 2:06
  • $\begingroup$ TL;DR -- I think that [clustering] $\leftarrow $ [segmentation] makes sense, but at a medium level of confidence. However, I don't think that [clustering] $\leftarrow $ [image-segmentation] makes sense. $\endgroup$ – Sycorax Nov 27 at 23:39
  • $\begingroup$ @Sycorax, I never meant to suggest [c] <- [is]. I think before the mapping is made, some of the threads currently tagged [s] should be retagged to [is]. I further think no good will come from the continued existence of the [s] tag. I'd suggest making the mapping, merging them, then deleting the mapping. That will make the [s] threads into [c] threads & [s] will completely disappear. $\endgroup$ – gung - Reinstate Monica Nov 28 at 1:48
  • $\begingroup$ @Sycorax: Then I think there are posts under [s] that is not [c], so for the moment I will not do the synonym, but clean up the wrong tags and try to write a better wiki. I also think that time series segmentation and clustering are very distinct, look at the wikipedia article. $\endgroup$ – kjetil b halvorsen 20 hours ago
  • $\begingroup$ @kjetilbhalvorsen I think we're in agreement about time series segmentation and time series clustering being distinct. $\endgroup$ – Sycorax 20 hours ago
  • $\begingroup$ @kjetilbhalvorsen, my suggestion, & what has been the typical protocol, is that the tag is cleaned up 1st & then the synonym is made. The fact that the tag needs to be cleaned up 1st isn't a reason not to make the synonym--it's almost always the case that the tag needs to be cleaned up 1st. $\endgroup$ – gung - Reinstate Monica 20 hours ago
  • $\begingroup$ @gung: Yes, but my point is that some of the [s] posts are not clustering at all, and since there is no other tag that can be used, the [s] tag must be left, with a better wiki and far fewer posts, but not synonymized. $\endgroup$ – kjetil b halvorsen 19 hours ago
  • $\begingroup$ @Sycorax: I did try some clarification here: stats.stackexchange.com/questions/74351/… $\endgroup$ – kjetil b halvorsen 18 hours ago

You must log in to answer this question.

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