13
$\begingroup$

How do we decide if a classifier is linear or non linear ?

What property/characteristic makes a classifier linear or non linear ?

Eg: Why SVM is a linear classifier ? Why Logistic Regression is linear classifier even though it uses logistic function which is a non linear function ?

$\endgroup$
3
  • 1
    $\begingroup$ The linearity of the classifier refers to its decision boundary. Is it a hyperplane or not? (And SVM's is generally not linear.) $\endgroup$ – Emre Oct 25 '15 at 1:36
  • $\begingroup$ Thanks @Emre , so as per your answer, if the feature space is say 6 dimensional and the decision boundary is 5 dimensional then it is a linear classifier. Am I right ? $\endgroup$ – mach Oct 25 '15 at 1:48
  • $\begingroup$ So it doesn't depend on the sixth one? It could still be nonlinear in the other five. $\endgroup$ – Emre Oct 25 '15 at 1:54
10
$\begingroup$

A classifier is linear if its decision boundary on the feature space is a linear function: positive and negative examples are separated by an hyperplane.

This is what a SVM does by definition without the use of the kernel trick.

Also logistic regression uses linear decision boundaries. Imagine you trained a logistic regression and obtained the coefficients $\beta_i$. You might want to classify a test record $\mathbf{x} =(x_1,\dots,x_k)$ if $P(\mathbf{x}) > 0.5$. Where the probability is obtained with your logistic regression by: $$P(\mathbf{x}) = \frac{1}{1+e^{-(\beta_0 + \beta_1 x_1 + \dots + \beta_k x_k)}}$$ If you work out the math you see that $P(\mathbf{x}) > 0.5$ defines a hyperplane on the feature space which separates positive from negative examples.

With $k$NN you don't have an hyperplane in general. Imagine some dense region of positive points. The decision boundary to classify test instances around those points will look like a curve - not a hyperplane.

$\endgroup$
2
  • $\begingroup$ Thanks ... but with a hyperplane we can classify into only two classes ( ie positive or negative) , then how do these classifiers (Logistic which is binary) handle multilabel classifications ? $\endgroup$ – mach Oct 25 '15 at 2:07
  • $\begingroup$ Yes hyperplanes deal with binary classification problems. There are two ways to use hyperplanes for multiclass classification: one vs one, and one vs all. Have a look at this wiki page: en.m.wikipedia.org/wiki/Multiclass_classification $\endgroup$ – Simone Oct 25 '15 at 2:14

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.