|
The classification result of Libsvm includes the confidence for positive class and negative class. I do not how is this confidence value calculated? I ran the Libsvm against a given data set, the maximum confidence value is around 0.7 and the minimum confidence value is around 0.2. What do these confidence value distributions tell us? For another data set, the maximum confidence value is almost 0.99, and the minimum confidence value is 0.01. Can I say this data set is more separable than the above data set? |
|
I assume that by confidence you mean the probability estimates, is that correct? If so, note that these estimates represent a function of the distance to the separating hyperplane, namely 1/(1+e^(A*f+B)) [http://www.csie.ntu.edu.tw/~cjlin/papers/svmprob/svmprob.pdf], so if the classes are clearly separable and there's some space in between, one may expect that fewer samples will have probability estimates in the vicinity of 0.5. The maximum and minimum of the probability estimate values have to do with the margin width and its relation to the maximum distance of data samples from the separating hyperplane, which has little to do with separability. |
|
I would like to know how you get the maximum confidence value ? And how we can use it as threshold for our system to get diffrent recall and prestion ? probability estimate is for each data samples , am lost could someone tell me the diffrent please |
