0
$\begingroup$

This question already has an answer here:

I have a 2-class problem involving many features. Does a linear support vector machine (SVM) classifier only take into account the values of these features and nothing more? Does it see the relationships between the variables? For example, does it work like this: if feature number 80 is "0", then feature 2: needs to be over 0.2 for class 1 and if feature number 80 is "1" then feature 2: needs to be over 0.8 to be in class 1? Does it do it with many variables at once, and do the values of other variables influence how the SVM is influenced by other variables?

$\endgroup$

marked as duplicate by gung, whuber Jan 21 '14 at 17:38

This question has been asked before and already has an answer. If those answers do not fully address your question, please ask a new question.

0
$\begingroup$

Those are good questions. Linear support vector machines look for the linearly-separating hyperplane in your data. The characteristics of that plane have a lot to do with the number of features in your classification problem, as well as the approach you're using to model the features (e.g., binary or term frequency/inverse document frequency, also known as tf-idf). It might help you to brush up on some SVM fundamentals. To this end, i've found "A Practical Guide to Support Vector Machine Classification" to be an outstanding reference. That said, SVM does use all the features you give it, but it is not necessarily equally influenced by all features. For example, in text classification, which tends to involve millions of features, it has been my experience that less-informative features, such as the presence or absence of a very common word, does not have as much influence as a less-common word that tends to be associated with one particular class.

$\endgroup$
  • $\begingroup$ The accepted answer for the "Help me understand SVM", pointed out by @whuber is probably a good place for you to start! $\endgroup$ – Kyle. Jan 21 '14 at 17:41

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