SVM Fitcsvm() thresholds vs ROC curve thresholds

I am using SVM for binary classification and the Mdl=fitcsvm() in MATLAB returns trained model Mdl containing info like Alpha, Bias, KernelParameters, etc. I test my testdata on this model over a range of threshold values to plot a ROC curve. But I have 2 questions in this whole process.

1. The Mdl never tells what threshold was used? Like >threshold is +ve class; <threshold is -ve class. Or is there a default threshold it uses? Is it 0?
2. In the ROC I'm plotting, let's say I decide on a threshold point that gives me satisfactory (TP,FP). But how do I use that? I mean the saved model Mdl does not surely use this. If default is 0 but I like 0.3 then how to make the Mdl know this when it is used in real application? Or should I explictly put in my code this 0.3? How? If it's impossible then what's the use of plotting roc curve in this context?

Answer 1

[the cyclist]

fitcsvm() fits the best possible SVM to the data. That function itself does not make the predictions.

The resulting model object from

Mdl=fitcsvm()

has a predict method that is used to make the actual predictions:

[class,score] = predict(Mdl,X)

For an SVM, the score output indicates a likelihood that a label comes from a particular class. A positive value means it is likely to be from that class. MATLAB will assign the class output accordingly.

If you are unsatisfied with that, then I believe you will need to calculate the posterior probability yourself, from the score. There is a section on classification scores and posterior probability in that documentation I linked. It admittedly looks a bit involved, but I don't know of a simpler way.

I see that there is a fitSVMPosterior function that looks like it might be useful in what you are doing. I also suggest searching on "scoreTransform", which I see popping up around different functions. I have used some of this a long time ago, but memory fades a bit.

I know that in scikit-learn, some models have a input parameter where you can ask for the output to be the posterior probability instead of the score, but I don't know for sure if there is an equivalent in MATLAB.