Normalization of unbalanced confusion matrix?

[hahong]

For dprime_from_confusion_ova()

Need to search literature to find out "standard" way of doing this.

[hahong]

d' is independent of the threshold (or decision boundary), so it is immune to unbalanced positive and negative samples --- i.e., no need to balance. Am I correct? cc: @npinto, @cadieu, @yamins81

[cadieu]

Correct, by definition, the d' considers the positive and negative
distributions independently and works with the mean and variance of each
distribution. So no need to balance.
But given only decisions (as in the case of human responses) you can also
compute d', and these decisions have some implicit threshold, which if
changed, could change the d'.

Also, here's that discussion of ROC vs. PrecisionRecall curves, which
relates to balancing the dataset (see bottom of page):
http://www.eecs.berkeley.edu/Research/Projects/CS/vision/bsds/

On Wed, Jul 18, 2012 at 1:06 PM, Ha Hong <
reply@reply.github.com

wrote:

d' is independent of the threshold (or decision boundary), so it is immune
to unbalanced positive and negative samples --- i.e., no need to balance.
Am I correct? cc: @npinto, @cadieu, @yamins81

---

Reply to this email directly or view it on GitHub:
#2 (comment)