How could I create a customised performance function for a neural network ?
3 views (last 30 days)
Show older comments
Hi!
I have imbalanced data to classify thus the mse performance function is not suitable, just as the mae, sae and sse.
So, I would create a new performance function based on sensibility and specificity but I have not found any way to edit it.
The only thing I found is "template_performance" but it's obsoleted for Matlab 2012 and, anyway, I don't understand how manage with it.
So, please, could you provide me with an example or a tutorial ?
Thanks by advance
0 Comments
Accepted Answer
Greg Heath
on 12 Jan 2013
I have never had reliable results with a MLP when the training priors differed by more than a factor of 2.
If you cannot oversample the underrepresented class, then undersample the overrepresented class (with each subsample no larger than twice the size of the smaller class).
A good way to subsample the larger class is to cluster it into multiple localized subsets that are subsequently randomly sampled.
Combine the results of independently trained multiple nets in either an ensemble (combine probability estimates) or a commitee (combine classification votes).
I have never had reliable results with a MLP using the noncontinuous misclassification error as a direct minimization goal.
Minimize MSE or weighted MSE for 0 or 1 targets.
Vary the MSE error weights until you can get approximately equal MSEs for both classes.
Use a holdout validation set and a varying threshold from 0 to 1 in order to get your operating curve.
Hope this helps.
Thank you for formally accepting my answer.
Greg
More Answers (4)
Greg Heath
on 5 Jan 2013
Edited: Greg Heath
on 5 Jan 2013
I have written about the unbalanced classification problem many times.
Try searching comp.ai.neural-nets and the CSSM newgroup
heath unbalanced
Stop laughing.
The quickest solution is to duplicate vectors in the smaller classes so that all classes have equal sizes.
Then, for c classes, use columns of the c-dimensional unit matrix as targets
Hope this helps.
Thank you for formally accepting my answer.
Greg
0 Comments
Charles Henri
on 11 Jan 2013
1 Comment
Omer
on 4 Feb 2014
I have also similar problem:
I am using nn toolbox functions to create a neural network for classification purpose (2 output neuron). Instead of using standard performance function to optimize, I want to use my own custom. Such that my performance function will be:
( fp/(fp+tn) ) + ( fn/(fn+tp) );
where
tp: true positive fn: false negative and so on. Of course output y of the network must be converted to 0 or 1. maybe like this:
yPred = ( y(:,1) > y(:,2) );
How can I do this with using newpr or newff?
any help appreciated Thanks
Greg Heath
on 12 Feb 2014
This may be of interest:
http://www.mathworks.com/matlabcentral/answers/56137-how-to-use-a-custom-transfer-function-in-neural-net-training
Greg
0 Comments
Greg Heath
on 13 Feb 2014
You have the misleading impression that unbalanced data requires changing the minimization objective function.
It does not.
If you duplicate some of the underrepresented class members and then modify them slightly by adding a little noise, the imbalance problem is solved.
You can then weight the posterior probability estimates with class-conditional prior probabilities and misclassification rates to forma risk function via Bayes Theory. The input is then assigned to the class that results in minimum risk.
This is classical pattern recognition covered in any decent pattern recognition text.
I have classified the BioID data set using this technique. Search the NEWSGROUP and comp.ai.neural-nets using combinations of search words like
greg BioID unbalanced priors
Hope this helps.
Greg
0 Comments
See Also
Products
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!