2
$\begingroup$

I have liblinear model file for a classifier learned using logistic regression. In the file, they say, the weight vector and intercept term. But when I simply plot it as $$w^Tx + b$$ on the original data, it does not seems like the proper decision boundary. But in case of SVM, plotting

$$w^Tx+b$$ seems to be the correct way. Is there any difference in case of Logistic Regression with liblinear ?

$\endgroup$
4
$\begingroup$

Logistic regression does not have decision boundaries. It is a method to estimate probabilities of events/class membership. Decisions are made in a separate step once you know the estimated risk along with utilities/costs/loss function, which is the way optimum decisions are made.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ but if i am not mistaken, it is among the linear classifiers, and the probability of class membership is decided based on $$\frac{1}{1+exp(-f(x))}$$ where $$f(x) = w^Tx+b$$, which is a linear decision boundary. $\endgroup$ – Shew Dec 4 '14 at 16:57
  • $\begingroup$ That would be true if "decision" were part of logistic regression. It is not. $\endgroup$ – Frank Harrell Dec 4 '14 at 16:58
  • $\begingroup$ what do you mean by "decision" being part of logistic regression ?. In normal settings (without any thresholding), the decision is similar to the case of SVMs, right ?. +ve, $$if p(x) \geq 0.5$$ (similar to $$f(x) >= 0$$ $\endgroup$ – Shew Dec 4 '14 at 17:03
  • $\begingroup$ I said that "decision" is not part of logistic regression. The construct you wrote which uses a highly arbitrary 0.5 risk threshold is not part of logistic regression. Logistic regression is a direct probability estimator. $\endgroup$ – Frank Harrell Dec 4 '14 at 17:10
  • $\begingroup$ Ok. In that case, what is the $$w$$ and $$b$$ in the logistic model file ?. What is the plane $$w^T+x$$ corresponds to ? $\endgroup$ – Shew Dec 4 '14 at 20:06

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.