1
$\begingroup$

Suppose I have a multi-class classification task. As far as I can tell, the primary metrics used for evaluating performance on this task is either to compute the confusion matrix, or the per-class f score.

However, I was wondering if we could apply cluster evaluation measures, such as the one proposed here, to serve as another metric for this multi-class classification task. If we put all datapoints with the same predicted class label into their own cluster, and compare the clusters induced by our predicted labels vs the clusters induced by the actual labels, would the resulting cluster evaluation metric also serve as a good classification metric?

$\endgroup$
2
  • $\begingroup$ V-measure is an example of external clustering criteria (= external clustering validation indices). Their difference with the classification performance indices is in that they require not to know class labels, that is, the one-to-one correspondence between classes of the predicted and classes of the actual partitions. (You might want to read this and all the formulas of both types of indices in my document 'Compare partitions' on my web-page.) $\endgroup$
    – ttnphns
    Feb 28, 2020 at 2:26
  • 1
    $\begingroup$ So, yes, sure, you may use the clustering indices instead of classification indices, to evaluate classification. it won't be illegal. But that means you simply disregard information partly, "forgetting" about that you know class labels. $\endgroup$
    – ttnphns
    Feb 28, 2020 at 2:31

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.