In Platt scaling[1], you fit a function to the predicted raw scores from a trained model, on a test set, in order to convert future scores into probabilities.

The function is: p(x) = 1 / (1 + exp(a*x + b))

x being a raw score. So, two parameters (a, b) are fit.

What I don't like in this method is that a model about the probabilities is assumed (a logistic curve here).

If Leo Breiman was still with us, what would have he used or invented?

Put another way: are there non parametric alternatives to Platt scaling?

Thanks for your input.

Bibliography: [1] Platt, J. (1999). Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in large margin classifiers, 10(3), 61-74.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.