37
$\begingroup$

I have a Machine Learning course this semester and the professor asked us to find a real-world problem and solve it by one of machine learning methods introduced in the class, as:

I am one of fans of stackoverflow and stackexchange and know database dumps of these websites are provided to the public because they are awesome! I hope I could find a good machine learning challenge about these databases and solve it.

My idea

One idea came to my mind is predicting tags for questions based on the entered words in question body. I think the Bayesian network is the right tool for learning tags for a question but need more research. Anyway, after learning phase when user finishes entering the question some tags should be suggested to him.

Please tell me:

I want to ask the stats community as experienced people about ML two questions:

  1. Do you think tag suggestion is at least a problem which has any chance to solve? Do you have any advice about it? I am a little worried because stackexchange does not implement such feature yet.

  2. Do you have any other/better idea for the ML project that is based on stackexchange database? I find it really hard to find something to learn from stackexchange databases.


Consideration about database errors: I would like to point that although the databases are huge and have many instances, they are not perfect and are prune to error. The obvious one is the age of users that is unreliable. Even selected tags for the question are not 100% correct. Anyway, we should consider the percent of correctness of data in selecting a problem.

Consideration about the problem itself: My project should not be about data-mining or something like this. It just should be an application of ML methods in real-world.

$\endgroup$
28
$\begingroup$

Yes, I think tag prediction is an interesting one and one for which you have a good shot at "success".

Below are some thoughts intended to potentially aid in brainstorming and further exploration of this topic. I think there are many potentially interesting directions that such a project could take. I would guess that a serious attempt at just one or two of the below would make for a more than adequate project and you're likely to come up with more interesting questions than those I've posed.

I'm going to take a very wide view as to what is considered machine learning. Undoubtedly some of my suggestions would be better classified as exploratory data analysis and more traditional statistical analysis. But, perhaps, it will help in some small way as you formulate your own interesting questions. You'll note, I try to address questions that I think would be interesting in terms of enhancing the functionality of the site. Of course, there are many other interesting questions as well that may not be that related to site friendliness.

  1. Basic descriptive analysis of user behavior: I'm guessing there is a very clear cyclic weekly pattern to user participation on this site. When does the site get the most traffic? What does the graph of user participation on the site look like, say, stratified by hour over the week? You'd want to adjust for potential changes in overall popularity of the site over time. This leads to the question, how has the site's popularity changed since inception? How does the participation of a "typical" user vary with time since joining? I'm guessing it ramps up pretty quickly at the start, then plateaus, and probably heads south after a few weeks or so of joining.
  2. Optimal submission of questions and answers: Getting insight on the first question seems to naturally lead to some more interesting (in an ML sense) questions. Say I have a question I need an answer to. If I want to maximize my probability of getting a response, when should I submit it? If I am responding to a question and I want to maximize my vote count, when should I submit my answer? Maybe the answers to these two are very different. How does this vary by the topic of the question (say, e.g., as defined by the associated tags)?
  3. Biclustering of users and topics: Which users are most alike in terms of their interests, again, perhaps as measured by tags? What topics are most similar according to which users participate? Can you come up with a nice visualization of these relationships? Offshoots of this would be to try to predict which user(s) is most likely to submit an answer to a particular question. (Imagine providing such technology to SE so that users could be notified of potentially interesting questions, not simply based on tags.)
  4. Clustering of answerers by behavior: It seems that there are a few different basic behavioral patterns regarding how answerers use this site. Can you come up with features and a clustering algorithm to cluster answerers according to their behavior. Are the clusters interpretable?
  5. Suggesting new tags: Can you come up with suggestions for new tags based on inferring topics from the questions and answers currently in the database. For example, I believe the tag [mixture-model] was recently added because someone noticed we were getting a bunch of related questions. But, it seems an information-retrieval approach should be able to extract such topics directly and potentially suggest them to moderators.
  6. Semisupervised learning of geographic locations: (This one may be a bit touchy from a privacy perspective.) Some users list where they are located. Others do not. Using usage patterns and potentially vocabulary, etc, can you put a geographic confidence region on the location of each user? Intuitively, it would seem that this would be (much) more accurate in terms of longitude than latitude.
  7. Automated flagging of possible duplicates and highly related questions: The site already has a similar sort of feature with the Related bar in the right margin. Finding nearly exact duplicates and suggesting them could be useful to the moderators. Doing this across sites in the SE community would seem to be new.
  8. Churn prediction and user retention: Using features from each user's history, can you predict the next time you expect to see them? Can you predict the probability they will return to the site conditional on how long they've been absent and features of their past behavior? This could be used, e.g., to try to notice when users are at risk of "churn" and engage them (say, via email) in an effort to retain them. A typical approach would shoot out an email after some fixed period of inactivity. But, each user is very different and there is lots of information about lots of users, so a more tailored approach could be developed.
$\endgroup$
  • 1
    $\begingroup$ @ cardinal. Thats a wonderful answer, and given the availability of all this data, it would make a fascinating project. $\endgroup$ – richiemorrisroe Apr 24 '11 at 15:40
  • 1
    $\begingroup$ Most of your suggestions seems practical and ML-related to me. Anyway, some of them should deal with erroneous and incomplete data. Sadly I don't have deep knowledge of data mining and cleansing nor enough time to learn it. I hope other members of stats do some work about these ideas and make a contribution to SE community and impress them :) $\endgroup$ – Isaac Apr 25 '11 at 6:27
  • 2
    $\begingroup$ @Isaac, the list I've provided was not intended to overwhelm. It was intended simply to potentially help with brainstorming. Depending on the nature of the project, I would think that handling 1-2 of them would be the most that could be expected. Cheers. $\endgroup$ – cardinal Apr 25 '11 at 22:56
  • 1
    $\begingroup$ all ideas are good or great, but I like "Biclustering of users and topics" best ... a recommender system for potentially interesting questions would be awesome. $\endgroup$ – steffen Apr 26 '11 at 7:12
9
$\begingroup$

I was thinking about tag prediction, too, I like the idea. I have the feeling that it is possible, but you may need to overcome many issues before you arrive to your final dataset. So I speculate the tag prediction may need a lot of time. In addition to incorrect tags the limit of max 5 tags may play a role. Also that some tags are subcategories of others (e.g. “multiple comparisons” can be viewed as a subcategory of “significance testing”).

I did not check if up-vote times are included in the downloadable database, but a more simple and still interesting project could be to predict the “final” number of votes (maybe after 5 months) on a question depending on the initial votes, and the timing of accepting an answer.

$\endgroup$
  • $\begingroup$ From what I seem to remember, for each user you have his/her vote with date + question ID. $\endgroup$ – chl Apr 23 '11 at 6:57
  • $\begingroup$ (+1) for vote prediction. Great idea! $\endgroup$ – steffen Apr 23 '11 at 7:19
  • 1
    $\begingroup$ This project seems great, especially if we predict the vpvote count very soon for the user. A further work can be to tell the user what is holding his/her question back and what improvement can make his question popular. Anyway, as always the choice of feature is really an important and challenging task and the performance of such predictions highly depends on this selection. TL;DR I like your idea $\endgroup$ – Isaac Apr 25 '11 at 6:32
2
$\begingroup$

This is a good question. I too have thought that the publicly available StackExchange datasets would make good subjects for analysis. These are sufficiently unusual that they might also be good testbeds for new statistical methods. Having such a large amount of well structured data is unusual, at any rate.

cardinal suggested a bunch of things which would actually be useful for StackExchange. I won't restrict myself to this.

Here is one obvious candidate for analysis, though it has no obvious use that comes to mind. It is a noticeable effect that high rep users are more likely to get upvotes, other things being equal. However, this effect is probably non-trivial to model. Since we can't compare usefulness across users very easily, an obvious approach would be to assume a users answers were always equally useful (not true in general but one has to start somewhere) and then add an inflationary term to account for his increasing reputation. One could then (I suppose) add in some terms that would account for his answers getting better with increasing experience. Maybe this could be handled by some kind of time series. I'm not sure how the data being interval would affect this. It might be an interesting exercise.

I'll add more examples if/when I think of them.

Is anyone aware of statistical research papers based on SE data? Also, Isaac mentioned that the data has errors. Does anyone know anything more about this?

$\endgroup$
  • $\begingroup$ This is, indeed, an interesting question and one that I believe AndyW began to analyze in a blog post and question awhile back. I do find curious your statement that there is a "noticeable" such effect, which I'm not entirely convinced actually exists. You then go on to suggest a way to model this, but doesn't that usually include making an attempt to answer the precise question you say already has an affirmative answer? $\endgroup$ – cardinal Nov 10 '11 at 17:51
  • $\begingroup$ @cardinal: Do you have a link to the blog post? I'm not sure I understand your last sentence. Tes, I think the effect is real and noticeable, at least if the user's name is Skeet, but at this point this is just an anecdotal impression, though a strong one. So, if you prefer, you can replace "noticeable effect" with hypothesis. The analysis would of course attempt to confirm or deny it, as well as measuring the strength of the effect if it existed. $\endgroup$ – Faheem Mitha Nov 10 '11 at 17:58
  • $\begingroup$ See the blog post, Does Jon Skeet have mental powers that make us upvote his answers? and related criticism. $\endgroup$ – Andy W Nov 10 '11 at 18:54

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.