How do I deal with different class sizes when classifying data with a petternnet?

I want to classify datasets using a patternnet. I have 2 classes (labelled 1 and 2). However, class 2 is significantly smaller than class 1 (ratio 1:9). The patternnet always classifies every sample into class 1, reaching 90% accuracy with it.

Is there any way to weigh or prioritize my classes so that this is not viewed as the best solution? (e.g. a cost matrix like for a decision tree (fitctree))

[MaHa]

Did you normalize your data prior to classification ? With normalisation (mapminmax function for example) you will not have a 1:9 ration anymore.

[Anne Bernhart]

I'm referring to the number of samples per class when I say smaller and 1:9 ratio. I don't see how normalizing my values could change that.

[MaHa]

I see I misunderstood sorry. What happens if you reduce the number of labbelled 1 to the number of labbelled 2 ? Does it still classes everything in L1 ?

[Anne Bernhart]

If I do that it works fine but I would prefer including all of my data.

Answer 1

[Shravan Kumar Vankaramoni]

Hi Anne,

Have a look at the below thread. Hope that answers your question

https://www.mathworks.com/matlabcentral/answers/501704-handling-imbalanced-data-with-patternnet

[Anne Bernhart]

thank you very much! That is exactly what I was looking for :)