The Spider Objects

 
 
 
 
   
  Probabilistic outputs for SVM
   
 


   A=PLATT(C,H) returns a platt object initialized with hyperparameters H and
               based on algorithm C. 
   Converts a real valued margin producing pattern recognition
   algorithm into a conditional probability estimator. It thus
   requires a pattern recognition approach that can produce a real
   valued output before thresholding, e.g SVMs. It achieves this
   by finding the coefficients of a sigmoid by using cross validation.
  
   Hyperparameters, and their defaults
  
    child = svm          -- real valued margin algorithm to be used 
    folds = 3            -- number of cv folds
  
   Model:
    A                -- 
    B                -- Parameters of the sigmoid: 1/(1+exp(A*f(x)+B)
    
   Methods:
    training, testing 
   
   Example:
   [rr,a]=train(platt(svm),gen(toy))
     rr.X  contains now the probability of being class 2
     predictions 
   [sign(rr.X-0.5),rr.Y]

Reference : Probabilistic Outputs for Support Vector Likelihood Methods Machines and Comparisons to Regularized

Author : John Platt

Link : http://research.microsoft.com/users/jplatt/SVMprob.ps.gz