19
$\begingroup$

This question seeking example datasets was recently closed as off-topic. Per the FAQ,

Questions about obtaining particular datasets are off-topic (they are too specialized).

As I interpret things though, that does not apply in this case. The FAQ seems concerned with questions like "Where can I find historical data about stock prices for the S&P 500?" or "Where I can find economic data about the GDP of every country in 2012?". Those questions are indeed about particular datasets and not really about data analysis, so those are the types of outlawed questions that I imagine when reading the FAQ. I believe those types were also the focus of these related Meta questions here and here.

However, the linked question is about data analysis. In particular, the asker is trying to develop new types of data analysis and asks for help in evaluating new algorithms. The asker is not looking for particular datasets but rather datasets with particular attributes as they relate to machine learning. Nothing in the FAQ forbids this.

There also seems to be a precedent that this type of question is actually fine. See here, here, and here for a few examples. Of course this is not a duplicate of those as this question has specific requirements (lots of attributes and suitable for regression).

So should this question be re-opened? I think it is useful, interesting, and within the scope and therefore should be re-opened. If it should not be re-opened, then the FAQ needs to be clarified and lots of other questions need to be closed as well.

$\endgroup$
5
  • 11
    $\begingroup$ +1 I upvoted this question because it is thoughtful, well formulated, and well researched. I am interested in hearing the community's opinions, because they can guide not only the ultimate status of the referenced question but also how all future questions about obtaining datasets are moderated. $\endgroup$
    – whuber Mod
    Oct 8, 2012 at 20:51
  • 7
    $\begingroup$ +1, I agree that this Q is thoughtful, well formulated, & well researched. In addition, I've always thought the type of Q at issue here, & those listed as "precedent[s]", should be on-topic. I do agree w/ people that Q's just looking for specific datasets (eg, GDP data) should be off-topic, though. $\endgroup$
    – gung - Reinstate Monica
    Oct 8, 2012 at 21:45
  • 5
    $\begingroup$ I think the block quote above is evidence that the question is off topic and the existence of precedent doesn't really change that. To quote @whuber's comment made in the Statistics jokes thread: "Some rules benefit from being ... bent ... once in a while. However, please don't use the existence of this thread to justify creating new ones that fall outside our guidelines unless you think there is a very good reason to do so!". I also don't quite see how the question is about data analysis, as you indicated. Just my 2¢... $\endgroup$
    – Macro
    Oct 9, 2012 at 13:32
  • 1
    $\begingroup$ @Macro How is the question about data analysis? Well, it's relevant to people developing new techniques for data analysis. Validating the efficacy of new algorithms is an important part of data analysis, and that tends to require data. Also, the question is useful across specific domains. It's quite useful to people developing new predictive analytics algorithms. $\endgroup$
    – Michael McGowan
    Oct 17, 2012 at 20:35
  • 2
    $\begingroup$ Ahh, the problem with voting on meta. Downvoted because while the question is well-phrased, the idea that it's supporting (list type questions) is one I disagree with (and votes on meta are used to show dissent). $\endgroup$
    – casperOne
    Oct 21, 2012 at 18:18

2 Answers 2

Reset to default
12
$\begingroup$

I think I've probably been the most vocal opponent of data-finding questions on CrossValidated, but this one seems fine to me. My main complaint has been that these questions need answers from subject-matter experts, not statistical experts, and are therefore simultaneously not very educational for CV users outside of that subject matter expertise, and more likely to get better responses from a subject-matter StackExchange site.

That complaint doesn't apply here. CV users from just about any field have a reasonable chance of being able to answer the question, and could provide additional pointers for the process of evaluating novel methods against known datasets. That's "statistical analysis, applied or theoretical", right?

Is it a very interesting question? No. But I'd vote to re-open if I had the rep.

$\endgroup$
9
  • 11
    $\begingroup$ Here's a thought experiment: if I walked into the local university's statistics department and asked the faculty and students for pointers to a good source for income distribution datasets, I think someone would have an answer. Conversely, if I asked for good datasets for evaluating classification/regression methods, I'd bet almost everyone would have an answer. $\endgroup$
    – Matt Parker
    Oct 12, 2012 at 21:08
  • 2
    $\begingroup$ Why is this off-topic? I think some very hard-problems in statistics of the present day lie in the high-dimensional setting. It would be more useful-to have a collection maintained on threads like this. Stat exchange is the place! The discussion here has Nothing about 'high-dimensional' datasets. its all empty on this direction and is more about regression/classification in the 'generic' setting and away from Tukey's EDA- view at data apart from his problem-solving view. $\endgroup$
    – hearse
    Oct 17, 2012 at 20:23
  • 2
    $\begingroup$ Matt, I think if you walked in a Stat department, nobody will have a clue about income distribution data sets. Someone in an economics department could... but I'd be really surprised to find statisticians knowledgeable about it. Most statistical audiences I talked to do not know what Gini index of income inequality is, unless there's somebody who worked specifically in government/social statistics. $\endgroup$
    – StasK
    Oct 18, 2012 at 3:22
  • 1
    $\begingroup$ Conversely, I was asked by faculty at our local department, "We have developed this cool methodology -- you work in this area [latent variable modeling], do you have a dataset we could apply this to?" -- which is exactly the backwards way of developing new statistical methods. My deep belief is that statistics should be driven by applications, not by desire to publish a hollow scholastic paper in Annals or JASA or JSPI. (This belief did cost me a tenure-track job, so don't listen to me.) $\endgroup$
    – StasK
    Oct 18, 2012 at 3:23
  • 1
    $\begingroup$ @StasK How then would you propose that developers of statistical software ensure that their algorithms are any good? $\endgroup$
    – Michael McGowan
    Oct 18, 2012 at 16:17
  • 1
    $\begingroup$ Michael, see stata-journal.com/article.html?article=pr0001. To ensure that your software does the right thing, you need to know the truth, and for none of the real data sets you do. So somebody looking for a data set is a red herring that somebody is after publishing a paper on a cute algorithm with no real applications. $\endgroup$
    – StasK
    Oct 18, 2012 at 18:40
  • $\begingroup$ This discussion is a good example of why I think the question should stand. @whuber, any thoughts on re-opening? $\endgroup$
    – Matt Parker
    Oct 18, 2012 at 18:55
  • $\begingroup$ I might be more sympathetic to the argument that "this belongs on subject matter expert" sites if it wasn't extremely difficult to get those sites going - but that's somewhat orthogonal to the problem here. I've voted to reopen. $\endgroup$
    – Fomite
    Dec 13, 2012 at 0:42
  • $\begingroup$ I have to agree with @Stask's comment that, in order to really "test" your method, you need simulated data, since you know the "truth". $\endgroup$
    – Macro
    Dec 14, 2012 at 16:59
1
$\begingroup$

The way this question is phrased, it's Not Constructive by the vast majority of the Stack Exchange sites.

From "Q&A is Hard, Let's Go Shopping":

The former question provides the path of least resistance: a laundry list of products I can buy without thinking about it too much. But that answer will only be valid for a year at best. The latter question may take some thinking, but its answer will be valid forever

The way this question is phrased, it is an example of the former question in the quote above, which is of the form:

What are/where can I find/what is the best/etc items of type <X>

These are problematic for a number of reasons:

  • The voting mechanism doesn't work well for these items. Since the initial question is overly broad, the upvotes are being used to compare apples to oranges.
  • It tends to attract a number of links to sources (with little accompanying them), which are not real answers; link-rot is a huge concern here.
  • People don't maintain the lists properly.

That said, it's been suggested on Meta Stack Overflow (the town hall for all of the Stack Exchange sites) many times that lists should be maintained somewhere else.

That doesn't mean list questions can't live here, but it's strongly encouraged that Stack Exchange sites have very strict guidelines around them so as to not have your site filled with reference lists instead of real questions and answers (remember, Stack Exchange sites are question and answer sites). We want other sites in Stack Exchange to benefit from the experiences that sites preceding them have experienced.

$\endgroup$
7
  • $\begingroup$ Note that I am not the author of the cited question; I merely authored the Meta question as someone interested in answers to that question. $\endgroup$
    – Michael McGowan
    Oct 21, 2012 at 22:50
  • $\begingroup$ @MichaelMcGowan Tweaked to reflect that. $\endgroup$
    – casperOne
    Oct 22, 2012 at 0:04
  • 1
    $\begingroup$ I think a couple of those concerns don't apply here. Unlike hardware/software recommendations, datasets don't really become obsolete - especially not for this purpose. For example, the famous Anderson iris dataset is still brought up all the time - even though it's almost 80 years old. $\endgroup$
    – Matt Parker
    Oct 22, 2012 at 20:30
  • $\begingroup$ I think link-rot may also be less of a problem, since the datasets suggested are likely to be government-issued, associated with a scholarly journal, or otherwise archived at a level beyond the usual blog post. The significance of votes in this context and answer maintenance are definitely pertinent concerns, though. $\endgroup$
    – Matt Parker
    Oct 22, 2012 at 20:31
  • $\begingroup$ @MattParker I don't disagree, but consider: where is the best place to get that data set? What about link rot, what if the site goes down? Those are the things that lead to the Not Constructive nature of the post, not the quality of what's referenced. $\endgroup$
    – casperOne
    Oct 22, 2012 at 20:32
  • $\begingroup$ @MattParker Regarding link-rot, I'd have to disagree, unless you're including the data in the post, it's a concern, no matter how reliable you think the link is. On all sites, in Stack Exchange's eyes, a link to Google is just as volatile to some guy running his domain from his laptop. $\endgroup$
    – casperOne
    Oct 22, 2012 at 20:33
  • $\begingroup$ In addition to the idea that they don't go obsolete, unlike hardware/software recommendations, there is inherently value in using a known quantity when developing new statistical software, self-learning, etc. Commonly used, known entities have widely known quirks, documented results, etc. To be frank, a data set shopping list is actually pretty useful. $\endgroup$
    – Fomite
    Dec 13, 2012 at 0:45

You must log in to answer this question.

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