CODEWITHSUNDEEP
machine-learningsvm
user2823269
user2823269

Reputation: 7

Accuracy metric for multi class classifier

What could be a good metric to represent the accuracy of a multi-class SVM classifier. As I have read, f_score would not make much sense if the number of classes goes above two.

Upvotes: 1

Views: 527

Answers (1)

lejlot
lejlot

Reputation: 66775

The main question is what is your problem about. Once you answer this question you can choose appropriate metric. For example, if your problem is to "minimize the number of incorrect answers" then accuracy is your metric (number of correct guesses above all guesses). If your problem is to "maximize the quality of classification inside each of the classes" then you should use balanced accuracy (or GMean) which might be seen either as a reweighted accuarcy where small classes get big weights, or as a mean (arithmetic for balanced accuracy and geometric of gmean). of accuracies of each class independently. In other words - there is no such thing as a correct metric for a multi class classification, neither is for a binary case - it follows directly from your particular problem and your objective.

Upvotes: 1

Related Questions