10
$\begingroup$

My name is Will Jagy. I advised a youngster yesterday to ask his question here, and he did, Understanding similarity sensitive hashing algorithm in AdaBoost so I suppose I am following up to see if he is in the correct place.

More troubling was this case, https://mathoverflow.net/questions/82499/what-is-the-step-by-step-procedure-for-classifying-a-new-data-point-with-an-rbf-k where some angry words were exchanged. This still bothers me; I do think that EMS made an error in progressing from stackoverflow.com to MO rather than here. Eventually Suvrit, who did not want to be involved, commented "recommend that you have a look at Section 7.4 of the book "Learning with Kernels" by Schölkopf and Smola" which was an answer, plus it turned out that EMS had that book in his office.

To the extent that I have a specific question in the latter case, it is: is the EMS question acceptable here, and would it have received an answer?

There was little to no Latex, I should be able to post the entire title and question here:

What is the step-by-step procedure for classifying a new data point with an RBF kernel SVM?

I've read a ton of tutorials on non-linear SVMs and understand the 'kernel trick' reasonably well. But one thing that is not explained is exactly what you do with some new data once you want to classify it. What do you compute the kernel function on? Suppose I have some data and that I extract a feature vector v. I have a list of coefficients for the higher-dimensional separating hyperplane (this is what the RBF kernel training determines). Now I need to take an inner product between that list of coefficients and something computed from the feature vector v. It's very unclear to me what that thing-computed-from-v is supposed to be.

Added: This non-programming question is partially motivated by wanting to reverse-engineer a function from scikits.learn, so I also asked on Stack Overflow. In tinkering with the scikits.learn code, I have gotten closer to an answer to my math question, so please also check that question too. It is linked here: < https://stackoverflow.com/questions/8360253/how-to-extract-info-from-scikits-learn-classifier-to-then-use-in-c-code >

oc.optimization-control fa.functional-analysis st.statistics

EMS http://people.seas.harvard.edu/~ely/

$\endgroup$
  • $\begingroup$ I have not been long enough on the site to respond if this question is acceptable or not. But, I am a bit surprised to see that none of the 'old timers'/mods have not responded. Could a mod or someone with a knowledge of the history of the site respond? $\endgroup$ – varty Dec 6 '11 at 15:16
  • 2
    $\begingroup$ @varty and Will, it is definitely on topic, so no worries in that regard. The question is not within my ken though, so I can not say what your chances are of receiving an answer. I believe a more thorough response should address the hostility on the MO site though, and how as a site we should try to handle such situations here on Cross Validated. $\endgroup$ – Andy W Dec 6 '11 at 15:50
  • 1
    $\begingroup$ @AndyW I fail to see the relevance of how MO treated the question. All we have to decide is whether such a question is acceptable or not irrespective of where it comes from and irrespective of how it was treated at the source. Of course, just as any other question, it may be unanswered for a wide variety of reasons (e.g., lack of focus/clarity, lack of experts etc). $\endgroup$ – varty Dec 6 '11 at 17:35
  • 1
    $\begingroup$ @varty, the issue that I was bringing up, as I see it, is the code of conduct for people whom ask questions is the same on here as it is at MO. The reason for the hostility at MO wasn't because of the question content, it was because of the ensuing comments that escalated into arguments. $\endgroup$ – Andy W Dec 7 '11 at 13:21
  • $\begingroup$ Thanks to everyone. I thought to send the AdaBoost question here after researching tags on your Main. For the curious, let me emphasize that some comments at the question were deleted by moderators. A fuller story, but perhaps still not everything, is at meta.mathoverflow.net/discussion/1233/… $\endgroup$ – Will Jagy Dec 7 '11 at 20:42
  • 2
    $\begingroup$ Will, I am very sorry that your student had a bad experience on MO. The moderators at CV do everything we can to avoid such experiences and our community has been very well behaved: we are used to questions that take some back-and-forth to formulate effectively; it's part of what makes statistics different from mathematics. $\endgroup$ – whuber Dec 9 '11 at 23:11
  • $\begingroup$ Thank you, whuber. I ought to clear up one point. The question with "RBF kernel SVM" happened first. One MO user suggested that the questioner post here, which was ignored. The next day, a young person asked the AdaBoost question on MO. I thought, why not look into this a little more, so I searched here for related questions here. What I found was encouraging, so the very first comment on the MO AdaBoost question was me saying "You might have a better experience on stats.stackexchange, here are some numbers of questions posted on related matters." Not my student...I am pleased he came here. $\endgroup$ – Will Jagy Dec 11 '11 at 23:42
7
$\begingroup$

The question is appropriate, but whether it would have received the answer, I cannot say, since you never know. In general this question is definitely a better fit here than in MO, since it is more "applied".

The problem with the question is that it can be answered only by researcher which is familiar with the terminology, although it seems that the author might really have problem with something simple. It always helps to make the question self-sufficient as possible. In this case if the exact procedure would have been outlined (something like: suppose we have the sample $X$, we use the model $M$, we seek to estimate the quantity $\theta$, with the following steps, and I do not get the idea in this exact step) then the question would not have been closed as too localised.

In general CrossValidated is less hostile place (in a sense that tolerance level is higher for poorly formulated questions) than MO, but looking at the comments and the attitude displayed by the author of the question, I suspect that the outcome might have been similar.

$\endgroup$
  • $\begingroup$ I think a discussion of a specific user's behavior should be preferably be avoided. If you agree then could you please edit out any mention of the user by name/id etc. $\endgroup$ – varty Dec 7 '11 at 16:41
  • $\begingroup$ Thank you for answering. Some of the comments were deleted by moderators, including the first one (which certainly qualified as "hostile") which complained about the use of acronyms without any explanation. In case anyone is curious, this case is still being debated on MO Meta, at meta.mathoverflow.net/discussion/1233/… $\endgroup$ – Will Jagy Dec 7 '11 at 20:35
  • $\begingroup$ varty and mpiktas, on MO the person posting the original question is referred to as "the OP," which is probably "the Original Post-er." $\endgroup$ – Will Jagy Dec 7 '11 at 20:48
  • 1
    $\begingroup$ I've edited my answer, but I do not understand the objection. The name of the author of the question was mentioned in the original question, even with the link to the author's webpage. Why then I cannot use this name in my answer? $\endgroup$ – mpiktas Dec 8 '11 at 3:21
  • 3
    $\begingroup$ The issue is this: we do not know the full story of what happened on MO. We do not know who provoked who etc. It is my opinion that we should treat every question on its merit irrespective of what happened at some other site. Thus, any reference to the OP's behavior at MO is irrelevant to whether the question per se is acceptable or not. Perhaps, I should have been clearer in my first comment- my preference is to avoid discussing the behavior of users at another site when considering if their question is acceptable here. Let us give the OP and the qn a fresh start. $\endgroup$ – varty Dec 8 '11 at 3:27

You must log in to answer this question.

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