0

I'm doing positive/negative classification of a imbalanced dataset. About 30% of the samples in the data are positive, and the rest are negative. With some tuning of parameters and classification algorithms, I've made an SVC (using the RBF kernel) with about 86% accuracy, 46% precision and 26% recall. These are all pretty good numbers, but for this particular dataset, I'm wanting the precision to be higher, even at the cost of recall. Basically, I want to err on the side of positive to improve precision.

How do you generally "bias" a classifier for a particular class?

Thanks in advance!

naiveai
  • 23
  • 7
  • Eshan: the standard ways to handle this problem are discussed in the thread I just linked – specifically, for a support vector classifier the default is to use class weights (which are implemented e.g. in LIBSVM). If you have more questions about it after looking into those a bit, please ask! – Danica Sep 26 '16 at 14:30
  • Sorry, as well as thanks, @Dougal! I promise I'll look more carefully next time I ask a question. – naiveai Sep 26 '16 at 14:37
  • No worries! It can be hard to find what you're looking for if you don't know the right terminology. – Danica Sep 26 '16 at 14:39

0 Answers0