0

I’m building a model using Linear SVM from the Scikit-learn package in Python. I have found that Linear SVM performs much better on my training set than Logistic Regression.

My question is, is there a way to interpret the coefficients from my linear SVM in a similar way as we would interpret logistic regression coefficients? I can output the coefficient of each feature in my model using coef_.

I’m trying to figure out if there is a way to measure how effective each feature in my dataset is in regards to predicting an outcome. For example, a completely hypothetical situation, using logistic regression to predict whether a student is in an honors class or not, given the student’s test scores. Using Logistic Regression it’s possible to calculate an output along the lines of “a one-unit increase in the test score leads to an expected 10% increase in odds of the student being in the honors class”. Is it possible to calculate something similar using Linear SVM?

As you can probably tell, I’m a complete newbie when it comes to this and any help would be greatly appreciated! Thanks

Yaris
  • 1
  • 1
  • Just out of curiosity, when you say "I have found that Linear SVM performs much better on my training set than Logistic Regression", are you sure they're using the exact same features? It's possible that your SVM is using a transformation of the original features ("kernel trick") behind the scenes... – Adrian Jul 15 '14 at 20:58
  • Hi Adrian, thanks for getting back to me. Good question, as far as I know I should be using the same features. When I output my coefficients, I get the same number of coefficients as features. I left the LinearSVC as default, no specific kernal selected. Could it still be transforming my features behind the scenes? – Yaris Jul 15 '14 at 21:05

0 Answers0