Neural Network error weights to reduce false positive

I have a classification scenario where two outputs differ significantly in importance. Type 1 errors, false positives, must be avoided. Type 2 errors, missed positives, are much less important. How can I structure my neural network to reflect this? Help train specifies EW can be: "a Nox1 cell array of scalar values defining relative network output importance"

Experimenting with EW = [0.1; 0.9] etc has not influenced the portion of false positives.

Answer 1

[Greg Heath]

The classic approach to pattern recognition is to minimize Bayesian risk which, for c classes, is a double sum over classes of products of something like ( See a pattern recognition text for accurate details)

a priori class probabilities         P(i)
input conditional class densities    p(j,x)
misclassification costs              C(i,j) or C(j,i)?

The message is you can choose the costs to bias the decisions any way you want.

I have many posts re this issue. Search

 classification costs

in both the NEWSGROUP and ANSWERS as well as comp.ai.neural-nets.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Ariel Liebman]

Hi Greg, this is the one I am trying to find your answers on comp.ai.neural-nets.