1
$\begingroup$

Good morning all. For a while now I have been thinking about the possibility of combining machine learning or more generally data science techniques to help this site handle some on the common issues (this is also hopefully be discussed in this thread) faced by the moderators and the site as a whole. A few months back, I notice a competition on kaggle (a data science competition website) to do retagging (I believe, I will dig out a link and edit accordingly) so I know this is something stackexchange have been looking into..

So I guess my questions are:

1) what are the common issues and possible problems that can be solved via a data science approach (for MSE in particular)?

I think the closing of questions can be predicted, by using some form of natural language processing. Another topic I know exists already is the detection of fraudulent account behaviour ( I am not too sure how smart that is)

2) if data science projects already exist at MSE where can one read about the existing problems and possible solutions

3) a more general question are these stackexchange websites open source in terms of being able to check solutions, as per question 2?

Also, I did a quick word search on meta for machine learning but found nothing relating to this question (saying that, I could very well be wrong)

Cheers, Rob

$\textbf{update}$ Here is the link with the kaggle challenge, as proposed by facebook recruitment. There is also an associated question on SE meta.

$\endgroup$
5
  • $\begingroup$ Most of the moderation and quality control tools use hard criteria applied to certain metrics, the details of which are kept confidential by SE. I don't believe there has been any advertised use of machine learning techniques for this, but whether they are actually doing this in the backend, you have to ask a developer. $\endgroup$ Sep 16, 2014 at 11:45
  • 2
    $\begingroup$ @WillieWong They are hard at work on it, witnessed by the latest SE podcast: Are we that predictable?. Quote: "We did some science and we threw a bunch of data into Vowpal Wabbit (not a typo) and built a predictor of question quality" $\endgroup$
    – user147263
    Sep 16, 2014 at 12:11
  • 1
    $\begingroup$ @Thursday: Thanks for bring that up. It is not clear to me whether that LQ predictor has been rolled out and if so whether it has been rolled out to all SE or just SO. Do you happen to know? $\endgroup$ Sep 16, 2014 at 12:16
  • 1
    $\begingroup$ @WillieWong It's still in the works, as a part of a massive SE quality improvement project. This particular bit of the project involves complete reworking of the First Posts review queue. We'll know it's on when the review looks different. (Also, Shog9 and David Fullerton posted two different proposed solutions in that thread, so who knows what actually gets implemented. Perhaps DF has the advantage of home field, being VP of Engineering.) $\endgroup$
    – user147263
    Sep 16, 2014 at 12:21
  • $\begingroup$ Thursday, I will checkout the podcast. @WillieWong apologies I have updated the post with the link, but the link was for a facebook challenge with the data from stack exchange check this SE thread on SE meta $\endgroup$
    – Chinny84
    Sep 16, 2014 at 18:06

1 Answer 1

3
$\begingroup$

what are the common issues and possible problems that can be solved via a data science approach

  1. Redesign the front page to show each users the questions of interest to them. SE is working on this already.
  2. Filter out questions that are likely to be low quality, so that they do not appear on the site until after review. SE is working on this too.
  3. Detect duplicate questions that are not marked as such.
  4. Find mistagged questions (you mentioned this, but I include it anyway, since this is not a solved problem).
  5. Find non-closed questions that would be closed if someone put them under review today. Same for closed questions that would be reopened.
  6. Detect useless greetings ("Good morning") and closing lines ("Cheers") with enough reliability to delete them automatically. Same for please help Help HELP HALPPP and can anyone this quick question please and thank you so very much appreciated.

are these stackexchange websites open source

SE uses many open source technologies for peripheral purposes like search and editing, but their internal data-handling routines are not open sourced. See Which tools and technologies are used to build the Stack Exchange Network?

The software running all Stack Exchange sites is the same; only a few parameters vary between sites. So, in terms of actual implementation your question is not specific to the Math site. Of course, your solution may be specific to this one.

In any case, I think the likelihood of SE actually using any user-submitted solutions in their codebase is low to none. You could implement it locally on a downloaded data dump, though.

$\endgroup$
6
  • $\begingroup$ Cheers for the open problems. Firstly, I like to address that I was not going to propose, or submit any solutions to the above or any other open problems, but this post was to gauge the level of machine learning that a site like this (with very clever individuals) might already have implemented :). This answer, has really got me thinking of how one would go about solving such problems such as the recommendation engine in the first problem..though this is probably done naturally through tag selection? $\endgroup$
    – Chinny84
    Sep 16, 2014 at 18:02
  • $\begingroup$ I think the 3. problem would be interesting to solve, whereby before a question is submitted by the OP then it is cross referenced or validated as non-duplicate and would save a lot of hassle, but this requires a lot of backend technology that may or may not be existence with the SE websites..But thank you for your post!! $\endgroup$
    – Chinny84
    Sep 16, 2014 at 18:02
  • $\begingroup$ Stack Overflow has been using a recommendation algorithm for its homepage since 2010. It is currently under redesign. $\endgroup$
    – user147263
    Sep 16, 2014 at 20:33
  • $\begingroup$ Thank you once again. I will close this off. One last point, most of your links are pointing to meta stack overflow, is this the main place to look for these updates? cheers $\endgroup$
    – Chinny84
    Sep 17, 2014 at 9:27
  • $\begingroup$ Generally, netword-wide feature changes are discussed and announced on Meta Stack Exchange, see Recent feature changes to Stack Exchange. But in recent months, machine learning approaches have been primarily discussed on Meta Stack Overflow, since this is the site most in need of greater automatization. $\endgroup$
    – user147263
    Sep 17, 2014 at 11:17
  • $\begingroup$ Cheers for the response. I will sign up there also. I would also like to take this time, to thank you for your contributions on MSE and meta in particular. (I promised I wouldn't get emotional ha) $\endgroup$
    – Chinny84
    Sep 17, 2014 at 11:32

You must log in to answer this question.

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