How is the "predicted_confidence" of classification determined for a prediction from the Aster DenseSVMPredictor function? And can the predicted_confidence be interpreted as a probability of the classification? We already searched in the paper that is referred to in the Aster Analytics Foundation Userguide (Hash-SVM: Scalable Kernel Machines for Large-Scale Visual Classification” by Yadong Mu,Gang Hua, Wei Fan, and Shih-Fu Chang ). We cannot find the answer to our question in this paper. Normally, probabilities from SVM are determined as in: "Probabilistic Outputs for Support Vector Machines and Comparisons to Regularized Likelihood Methods, Platt (1999)".
The predict_confidence is a fraction between 0 and 1, which estimates the quality of the predict result. It is given by the following formula: 1/(1 + exp(-v)), where v is the internal predicted value.
If the predict_confidence is larger, the corresponding result will be more dependable.
Thanks for your reply. We still have some questions, hopefully you will be able to answer them as wel.
So the predict_confidence is not done by Platt scaling ( p(y|X) = 1 / (1 + exp(A * f(X) + B)) ) as done in the aforementioned paper?
Is there a way to extract the internal prediction?
And is the internal prediction equal to the signed distance of a sample from the hyperplane?
I am sorry. I do not have these details. In case you are a Teradata or Aster customer, you can open an incident. Customer support will try reaching out to engineering to gather these details. I am not sure whether all answers can be provided but it is worth a try.