CODEWITHSUNDEEP
binaryclassificationsvmprobabilitylibsvm
hollycrab
hollycrab

Reputation: 325

LIBSVM probability estimates

I would like to ask if it is true that the probability estimates produced by LIBSVM actually indicate the distance to the separating plane ?

If I have binary class, can I say that the smaller probability estimates of class 0 also means higher possibility of it belong to class 1 compare to bigger probability estimates of class 0 ?

Thanks

Upvotes: 1

Views: 1316

Answers (1)

alexandrekow
alexandrekow

Reputation: 1937

First question:

If the distance between the separating hyperplane and a vector is very large, we can have a high degree of confidence that it was correctly classified, thus the probability estimate will be high.

In contrary, if you have a vector very close to your hyperplane separating two classes, its probability estimate will be close to 0.5 for each class.

Keep in mind that the probability estimate is NOT the distance. But I would say it is computed by using the distance.

Second question:

Indeed if you have a binary class, the probability estimate will be spread between the two classes.

The event A "the vector belongs to class 0" then the event "the vector belongs to class 1" is NOT A

If the vector is in class 0, we compute

enter image description here

And we can compute the probability of NOT A

enter image description here

So if class 1 have a 0.8 probability, class 0 will have a 0.2 probability.

Upvotes: 4

Related Questions