1
$\begingroup$

I am trying to analyse my data before doing multi-class classification with SVM.

I have several variables. I pick one of them and study it.

This is a categorical variable. It can have the value 0 or the value 1.

Most of the time, it has the value 0. But sometimes it gets a 1.

I studied the frequency of ones and I got the following results:

The variable has the value 1 :

  • 2.71% for all classes
  • 2.22% for class 1.
  • 2.53% for class 2.
  • 4.79% for class 3.
  • 2.23% for class 4.
  • 1.99% for class 5.

Can I consider that this variable is important for predicting class 3 or is it incorrect?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
1
$\begingroup$

The immediate answer is: it's not clear yet - this analysis doesn't tell you enough to say.

You seem to want to say something like "I just picked up one data point, and it was a 1 - should I guess that it's class 3?" The first problem with that is that it depends on how MANY class 3 samples there are. If there were equal numbers of every category, it would be the best guess. But if there is only one class 3 sample, and a million from the other categories, it's a terrible guess!

The other thing to mention is that it's impossible to (meaningfully) classify your data into six categories using only one binary feature, so if you have other variables, you should see what else they tell you. They might have a very different story.

Improve this answer
$\endgroup$
3
  • $\begingroup$ Is there a reason you haven't run the svm yet? Are you looking for a different way to analyze things, or were you just curious about this approach? $\endgroup$
    – one_observation
    Commented Jan 4, 2016 at 21:56
  • $\begingroup$ I already ran the SVM but I am trying to learn how to do feature-selection. And it seems to me that a lot of it comes from analysing the data. But I don't really know what to look for as I am not very strong on statistics. I also wanted to see if I could find a specific variable which would be good for classifying a particular classe. So if I have 100 variables, I could pick 20 for classifying one class, 34 for classifying another and so on... $\endgroup$
    – Octoplus
    Commented Jan 4, 2016 at 22:14
  • $\begingroup$ Well, using an SVM is analyzing the data -- the point is understanding what you're seeing. Feature selection means finding only the features that matter, so in general, you can take features out, re-run the model, and see if your predictions are any worse. If not, you can remove that feature. There are more sophisticated ways of doing it, but that's the general principal. Also, I'm not sure exactly what you're suggesting by having subsets of variables for different classes, but in general, you only want 1 model for all your predictions. I suggest a tutorial on feature selection. $\endgroup$
    – one_observation
    Commented Jan 5, 2016 at 4:19

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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