# Calibration of SVM Probability

I am using 2-class SVM from e1071 R-Package. I have been recommended to calibrate the output of classifier. In this regard, I found this page. The code works for me and I want to apply it in my project; however, I am not sure about the vector of pred which its definition is "vector of predictions of each observation from the classifier. Should be real number."

The output of my classifier is like this:

    Decision values:
t34a t34b
0    0
attr(,"decision.values")
1/0
t34a -0.2859799
t34b -0.4527416
attr(,"probabilities")
1         0
t34a 0.17057557 0.8294244
t34b 0.07377427 0.9262257
Levels: 0 1


Do you know I should use which of this information for pred? decision values? or probability values of target class, say class 1, i.e. first column of probability values?

Also, do you know how I should apply it in learning procedure? Typically, the learning procedure is as follows:

1) estimation of performance of SVM using Leave-One-Out cross-validation.

2) training SVM with whole training samples,

3) and then evaluating the model by testing samples.

If I want to use calibration technique, where I can do that?

I appreciated if you could help me.