-1
$\begingroup$

Edit (for completeness)

The discussion was about a question I asked that was put on hold. The reasons for closure and what is needed to reopen the question is answered in the many comments below. Maybe I'll get back to it some time, but for the moment I've decided to delete the question because editing it would cause the question to bump to top which could lead to new discussion and I would like to avoid after this big meta. Thanks to everyone to invest so much time in this.

$\endgroup$
  • 6
    $\begingroup$ There's a simple principle here: it is not sufficient (or even necessary) that you view your own question as clear; it has to seem that way to others. Beyond that, what can happen is that if a moderator thinks your question has improved; then they can reopen it; or users with sufficiently high reputation can vote for it to be reopened. I just read your question and don't see anything forbidden: you would be told if that were so. But I can't see that your question is as yet precise enough to be a good question, so I won't vote to reopen. Others may have different views. $\endgroup$ – Nick Cox Aug 14 at 15:42
  • 4
    $\begingroup$ I agree with @Nick. Although the edits have helped, it's unclear to me what the relationship might be between your scores and predicting what you "like." If, for instance (and this might not be the interpretation you intend) your purpose is to estimate the chance that your score might be 8, 9, or 10, then stating that explicitly in the question would make it sufficiently clear IMHO. But currently, many other interpretations are possible, such as estimating your actual score. Suggesting, as you have, that we're looking only for non-"trivial" questions, is both wrong and unconstructive. $\endgroup$ – whuber Aug 14 at 15:57
  • 1
    $\begingroup$ @whuber I do not understand your different views on how my question can be interpreted. To me it is clear since I said that R2 decides what model is "best". So we have my preference, coded in the variable me and I am asking for another variable with predicted values. The agreement between those two variables in terms of R2 shall be high. How can I precise this further? $\endgroup$ – machine Aug 14 at 17:06
  • $\begingroup$ @whuber I was sure that it is because you find the question trivial. If not, I am sorry. To me it is unclear why the question is not clear enough. Maybe you can clarify that $\endgroup$ – machine Aug 14 at 17:08
  • 1
    $\begingroup$ But $R^2$ of what model, exactly? Here are some different ones for you to contemplate: (1) run a logistic regression where the response is $1$ for all scores of 8, 9, or 10 and otherwise is $0;$ (2) run an ordinal logistic regression where the response is the nominal score; (3) run an OLS regression where the response is the score viewed numerically; (4) run an OLS regression on a transformed score value (with the transform selected to achieve a near-linear relationship); (5) (I could go on and on...). The point is that what you "like" isn't anything one can model: be more specific. $\endgroup$ – whuber Aug 14 at 18:54
  • 1
    $\begingroup$ @whuber That is excatly my question: what excact model is the best (in terms of R2). So yes, it is correct, that I do not know wheter go for transformation or not and if yes, for which one - this is excactly what the question about (what to do to achieve high R2). It is unusual for question to state all this. If one knows all that - what would one ask for? $\endgroup$ – machine Aug 14 at 21:25
  • 1
    $\begingroup$ @whuber Mostly the questions are in same format as mine: "my data is.. What model to use to test significance?". In my case "my data is... What model to use for prediction?" $\endgroup$ – machine Aug 14 at 21:26
  • 2
    $\begingroup$ That is not what your question asks. In its present form it asks "What movies do I like." All I have requested from you is that you help us understand how your "score" might be related to this vague, unquantitative sense of "like" so that we won't waste your time and ours providing different, conflicting suggestions based on different interpretations of what you have in mind. (In fact, I don't even know whether your score of 1 would be liking or hating a movie!) I would be glad to close any question that is equally vague. That's enough discussion from me: you know where I stand. $\endgroup$ – whuber Aug 14 at 21:52
  • $\begingroup$ @whuber: alright, that is a precise question you ask. The scale is from 1= worst to 10= best rating. I will update the question. But in fact, this information isn't necessary to answer the question because if a prediction is good (in terms of R2 or LAD) then my question is answered and that is regardless whether 1 is worst or 10 is worst $\endgroup$ – machine Aug 15 at 5:36
5
$\begingroup$

"What would be the best way to predict the movies I like?" is arguably too broad (perhaps not, it's ultimately for the community to decide, but my guess from my experience on CV is that it might be). There will be lots of possibilities, and "best" is not well defined.

On the other hand, I don't think the level of "triviality" matters much for the site. There is no need for your project to be about death. Predicting movie preference is fine, and in fact, there has been a lot of academic work on more or less that topic within machine learning (e.g., recommender systems: CV, Wikipedia).

What might help here is to familiarize yourself with SE's processes for closing and reopening threads. Our help center has articles on closing, reopening, and how to ask a good question (you might also want to read our meta.CV thread on How to ask a “good” question on CrossValidated?). Relative to this conversation, a brief version is that a question can be put on hold by votes from 5 members of the community or from a moderator, if there are some concerns about the question. At that point, you can edit the question to address those concerns, and communicate with members in the comments. Editing your post bumps it into the reopen queue where members can view the issues and subsequent changes; they may vote to reopen the question or to leave it closed. If enough people vote one way or another, the issue is resolved (it is reopened or left closed). Notably however, this process takes time. You need 5 reopen votes from members, or 1 from a moderator, for your question to be reopened. On smaller sites like CV, this can often take a couple days. I recognize that is frustrating for you, but it is the way the system works.


Update
In a comment below, you write:

"There will be lots of possibilities" - yes, this is true! And that was something I was really looking for, to read different approaches and see what works best (again, best in terms of R2). You all probably know kaggle, where it is also about to find the "best model" in terms of some goodness or lack of fit. To me, my question appears to be similar. And well, nobody shuts down kaggle or puts in on hold for being to unclear. I know I can't compare kaggle with CV, just writing it to explain what I was expecting. I hoped for different approaches

This is in fact something I was wondering about. It seemed like maybe you had a kaggle-type model in mind, where many people would try many things and post them, and you would pick the one that you liked best (for example, maybe because it gave you the best $R^2$ on a hold out sample you had). That is very much not what CV is about. CV is not kaggle. This is a site for questions about statistics (machine learning, etc.), not a place for people to analyze actual data (although that does happen on occasion in the course of answering a question about statistics). That makes your question off topic here, irrespective of any triviality or lack thereof.

$\endgroup$
  • $\begingroup$ You write that "best is not well defined" but I did define "best" prediction in terms of R2. Further, in my opinion my question fulfills what the link says about a "good" question. If not, I would be happy to know what is missing. And ths is my problem: I do not know how to further edit it since @whuber does not reply to this $\endgroup$ – machine Aug 14 at 16:13
  • 1
    $\begingroup$ Not so; @whuber has commented above under your question. And, as already pointed out, we note your opinion but it really doesn't decide this. $\endgroup$ – Nick Cox Aug 14 at 16:32
  • 2
    $\begingroup$ I am sure your frustration is widely appreciated, but no one can suggest how to reword your question if they are unclear what it is. $\endgroup$ – Nick Cox Aug 14 at 16:47
  • 5
    $\begingroup$ @schwantke, I'll strike the part about best from the answer, if that helps you--it's a very small point. I have not voted either way (to close or reopen, nor voted on the Q itself), but I continue to suspect your Q will be viewed as too broad by the community (it's ultimately their decision). The main point of my answer here is that you shouldn't expect your post to be reopened immediately even if your edits resolve all concerns. It typically takes a while to work through the system. In addition, the potential triviality is irrelevant. $\endgroup$ – gung Aug 14 at 17:18
  • 5
    $\begingroup$ So, is this an answer to your original question?. Try all the models you regard as possible. Now choose that with highest $R^2$. If so, then well and good. If not, it seems that you're inviting wider comments and it's hard to know what fraction of any regression text should be part of an answer. It's also hard to tell from your description whether plain regression is even sensible. Rating is as I understand bounded and it's not clear whether it is best considered as measured or just ordinal. And so on. $\endgroup$ – Nick Cox Aug 14 at 17:28
  • 4
    $\begingroup$ I've heard of Kaggle but have no idea how it operates. Your profile shows that you are not exactlyl new to SE, so I am surprised you seem surprised at the reaction to your original question. But I am left bemused. If you define the best model as having the highest $R^2$ then what is there to ask about? That's a rhetorical question, as I am bailing out here. Saying repeatedly that I am not clear quite what you're seeking is not fun for me or constructive for you. Also, I don't agree that the best model is that with highest $R^2$. $\endgroup$ – Nick Cox Aug 14 at 18:19
  • 2
    $\begingroup$ @schwantke, hmmm, this was something I wondered about. It seemed like maybe you had a kaggle-type model in mind, where people would try many things & post them, & you would pick the one that you liked best (eg, maybe because it gave you the best R2 on a hold out sample you had). That is very much not what CV is about. That makes your question off topic here, irrespective of any triviality or lack thereof. $\endgroup$ – gung Aug 14 at 18:37
  • 1
    $\begingroup$ @schwantke, it's up to you. If it remains closed indefinitely, SE's bot will eventually delete it automatically, although that will take time. Regarding other questions, it is certainly the case that off topic Q's just slip though the system every so often. That isn't a justification for leaving open / opening another off topic Q. If you see one, you can flag it. If you're referring to this, it does not seem to envision a kaggle-like event where people analyze the data for the OP & submit results, nor does it seem obviously off topic to me. $\endgroup$ – gung Aug 14 at 19:16
  • 1
    $\begingroup$ @gung I probably should not have mentioned the comparison to kaggle. Yes, I was refering to that question which is "How to choose Hypothesis Test?". My question is just slighly different: "How to choose prediction model?". For both questions there different solutions based on what assumptions are made, how the data is, and so on. And I do not consider the question off-topic either. As I said I did not try to find a bad question. Another bad question does not justify anything. I consider it a normal question and compare mine to that one. $\endgroup$ – machine Aug 14 at 19:21
  • 1
    $\begingroup$ This is Meta and the focus should be about answering your question, at the top of the thread. You should be trying to revise the original question, as nothing else will help your case, and you are doing that. Sorry, but I don't think Meta comments are an appropriate place for me to discuss $R^2$, its use as a measure of model merit, etc. My Meta point was that your original question was raising too many different issues, being about how to analyse your data while also assuming that best $R^2$ was the way to work. Now the data are given and all of a sudden best LAD is the criterion. $\endgroup$ – Nick Cox Aug 15 at 7:47
  • 1
    $\begingroup$ @NickCox "your original question was raising too many different issues, being about how to analyse your data while also assuming that best R2 was the way to work.". How can there be any question? Whatever measure, R2, LAD or highest AUC in a ROC analysis: that are all well-defined measures and I ask for a model that maximizes them (or in case of error, minimizes) $\endgroup$ – machine Aug 15 at 10:07
  • 1
    $\begingroup$ @NickCox Regardless how hard I try I do not understand a little that it is not clear what I mean. There are thre variables. All I ask for is some model that can best predict on of the variables with the both others, best being measured in low LAD. What is unclear about that? $\endgroup$ – machine Aug 15 at 10:10
  • 1
    $\begingroup$ @NickCox "LAD is not a criterion I would dream of using for such data". About every criteria there can be critic. Here LAD fits best because in the end I want predictions that are (absolutely) close to my ratings. Why would I want to build strange categories (as suggested by whuber) or square the deviance? Therefore.. $\endgroup$ – machine Aug 15 at 10:32
  • 2
    $\begingroup$ Please consider me signed out of this discussion. I don't want to seem rude, or unconstructive, but for the third and really final time it is neither a good idea nor likely to be helpful to try to switch the debate you might be having on main about how to model your data to comments in Meta. There is still a hope for you that someone will vote to reopen. $\endgroup$ – Nick Cox Aug 15 at 10:33
  • 3
    $\begingroup$ Unfortunately that isn't clear, @schwantke. I still think the question is not a good fit for this site & should be closed. It has already accrued a close vote from the community, and I may step in if there are more. $\endgroup$ – gung Aug 15 at 15:51
4
$\begingroup$

a client asks the very same question and provides you the same data. Can you imagine asking the client "do you want to run an OLS regression on a transformed score value/ run an ordinal logistic regression...."

No; we can not expect the client to know all such details.

But at the same time clients are typically asking very open questions which should not be answered with a direct solution. (Such questions are considered to be bad questions, or at least questions not to be asked on this, I hope, technical website, working at a higher level than naive clients.)

I can not imagine that a statistician will answer a client directly with advice to use OLS regression on a transformed score value, run an ordinal logistic, etc. without first asking the client for further information what all the data is about and to get a better idea of the context and goals of the problem/question.

Thus, I would argue that my question is in the typical format "I have data XX. What model (for significance testing/ prediction/ ..) do I need to use?".

This format is exactly the problem. It is too broad. We have no idea what direction you would like to go. In this aspect questions that are asking for a model to accompany data are notoriously difficult. They are not questions that are guided by having a problem, but instead questions that are guided by having data. In the most stereotypical way they are questions like "I have data; can you tell me what to do with it?".

Personally, I imagine the data from your question might be explored by doing some kriging to generate a contour plot of your preference/rating as a function of the two other ratings. But whether this is a solution, or how to continue after exploring the data, I have no idea. It is unclear what you want. For example, if the model/question would be turned into an app, something practical, then what would the model be supposed to do or help out with? Would you expect something like an app that provides advice on how you would rate the movie based on the mean ratings of two others? (framing the question into a problem case, what an app that has your problem is supposed to do, might help you explain what the question/goal is really about)

I can understand that you get frustrated when your question gets rejected (and not the least at the discussion that follows). But please understand that the content on this website is created completely voluntarily. People who answer questions on this website are not doing this typically in a role as 'a statistics consultant helping a client'.

$\endgroup$
  • 1
    $\begingroup$ You write a statician would "without first asking the client for further information". Well, I that is something that would be great if it actually happened., because it would actually help to rephrase my question. Instead, I was provided links how to ask question and I saw literally no point I was violating or people kept telling the question is too broad, which does not help to rephrase, either. $\endgroup$ – machine Aug 16 at 5:58
  • 1
    $\begingroup$ "It is unclear what you want. (ie. if it would be turned into an app,(...)". This is a question ""for further information" which is nice. But my question is: Why is this relevant? if I want to use the prediction in an app, will people suggest another prediction model than if I use this in R? I do not see why this is relevant, thus I do not think that adding this information will lead people to reopen the question. Hence, I still do not know what exactly to do. $\endgroup$ – machine Aug 16 at 6:02
  • 1
    $\begingroup$ "The people that answer questions on this website are not doing this typically in a role as 'a statistics consultant helping a client'." I am aware of that. I have also a job that I do and I use SE quite a lot (although I'm usually on SO). But I am not to sure how this is related to the question what makes the question a bad one. $\endgroup$ – machine Aug 16 at 6:06
  • 1
    $\begingroup$ "This format is exactly the problem.". Well, this is a really typical CV-format. So maybe there is some need to discussion whether this fomat is accepted in general or not. If not, this can be added to FAQ and such question can be closed. But leaving soo many open (and often answering them) and now, for this one certain qestion, saying that the format is the problem and this is why the question should be closed, appear inconsistent to me. $\endgroup$ – machine Aug 16 at 6:11
  • 1
    $\begingroup$ "Why is this relevant? if I want to use the prediction in an app, will people suggest another prediction model than if I use this in R?" I would say that context is extremely important in statistics. $\endgroup$ – Sextus Empiricus Aug 16 at 6:28
  • 4
    $\begingroup$ "This format is exactly the problem.". Well, this is a really typical CV-format. ----- Data guided questions are not the intention of CV. The questions must be problem guided. Yes, when data is available then it can be added to the question, but only as an aid in order to either understand the question/problem (context) or to help to explain the answer. The CV site is not a place to help you look for problems (that go with your data) but instead a place to help you solve those problems. $\endgroup$ – Sextus Empiricus Aug 16 at 6:35
  • 2
    $\begingroup$ Regarding context; Note that there are many different settings and methods all of which aim to reduce R-squared. So context is necessary in order to deduce what sort of method is relevant. In addition you have not defined R-squared very well (e.g. is it defined for some validation set, which means optimize R-squared of predictions, or is it defined for the training set, which means optimize R-squared for the fit of the present data). Also confusing is that you have not addressed whuber's comment. You ask for the 'best'/'high R2', but there's no explanation for what model it needs to be done. $\endgroup$ – Sextus Empiricus Aug 16 at 7:18
  • 3
    $\begingroup$ While I suggest ways to improve ways to improve your question (like explaining what R^2 you wish to improve) Note that, at the same time, your problem might be difficult to 'salvage'. There is a set of problems that is not a good fit for CV. That is those problems which have both a more broader set of solutions (many ways to solve it, and many opinions how to solve it) as well as problems that are more specific to a single case (not well defined such that it can be easily linked to other people with the same problem, or at least not easy to be searched). $\endgroup$ – Sextus Empiricus Aug 16 at 7:43
  • 2
    $\begingroup$ I guess that a bit more context might help to, at least, understand a bit more your ideas behind the question, and (not the least) fill in the gaps of our understanding what you are trying to do. Context is not only helping to define the problem, but also helping to understand the language and phrases by which somebody tries to explain the problem (see Wittgenstein's brown and green books, although I realize while typing this, that it makes this conversation very offtopic and philosophical). $\endgroup$ – Sextus Empiricus Aug 16 at 7:46
  • 2
    $\begingroup$ Regarding feeling bitter about a question being closed/rejected: Sometimes an answer on SE is not provided in the form of an answer. The criticism on the question can be, in a way, some sort of answer as well. $\endgroup$ – Sextus Empiricus Aug 16 at 8:40
  • 2
    $\begingroup$ @schwantke No one takes the slightest pleasure in closing your questions or in your disappointment or your dissatisfaction with the experience. But CV is not a helpline; there is no entitlement to an answer if the community closes a question. Not constructive, sure, but the Catch-22 is here: Too broad, so you should ask something more specific; we can't choose for you. Unclear, meaning we don't see a clear question; there is no point in our guessing what that question might be. $\endgroup$ – Nick Cox Aug 16 at 9:16
  • 3
    $\begingroup$ @schwantke, it's not so nice to refer, in this way, to some better constructive feedback (which is implying other's couldn't). The way that I classify the discussion is mostly as a misunderstanding. From your side there has been just as much nonconstructive commenting which you should not (indirectly) blame on others by pointing out 'see, this person is able to understand me and comment the way I like'. When I read the comments and chat I see several good comments that just did not fell into place. These (good) comments should not be classified as nonconstructive (they just did not work well). $\endgroup$ – Sextus Empiricus Aug 16 at 10:58
  • 3
    $\begingroup$ I could summarize this discussion by quoting your 2nd comment: "This is a question ""for further information" which is nice. But my question is: Why is this relevant?" This way of replying to comments, by answering a question with a question, leads to a long winded (+ annoying) discussion. You could've lost me there. Such style (challenging feedback) gives the impression that you are not on the same wavelength (or trying to be) while the other people are mostly trying to help out with your question. Trying to defend/battle a question against closing is often not a good strategy here. $\endgroup$ – Sextus Empiricus Aug 16 at 11:12
  • 2
    $\begingroup$ You do realize that your question is unclear? That is, it is unclear to other people. It might be clear to you, but it is not clear to others (and you need to solve that discrepancy or otherwise there is no way to move forward). So, when the other people (that are trying to help you with your question) are asking for context or to fill in some gaps (like Fr1 and whuber had lots of questions that you did not address), then asking why all that is necessary is not helping out the people that are trying to help you. $\endgroup$ – Sextus Empiricus Aug 16 at 14:41
  • 3
    $\begingroup$ "..because I think adding this information will not reopen the question..." I wrote that this avoidance to answer questions and not following up with more information is typical for the discussion. Sure, I get it that now you feel less interested to change the question which is a bit loaded now. But for the sake of this meta question "why is my question still on hold?" my story about your question lacking context and being data-driven rather than problem-driven is just what it is. If you go against it ok, but then the discussion ends, I can not make it different why your question is on hold. $\endgroup$ – Sextus Empiricus Aug 16 at 19:43

You must log in to answer this question.

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