BAYESCLS |
Bayesian classifier with reject option.
Synopsis:
[y, dfce] = bayescls(X,model)
Description:
This function implements the classifier minimizing the Bayesian risk
with 0/1-loss function. It corresponds to the minimization of
probability of misclassification. The input vectors X are classified
into classes with the highest a posterior probabilities computed from
given model.
The model contains parameters of conditional class probabilities
in model.Pclass [cell 1 x num_classes] and a priory probabilities
in model.Prior [1 x num_classes].
The function
p = feval(model.Pclass{i}.fun, X, model.pclass{i})
is called to evaluate the i-the class conditional probability of X.
It returns class labels y [1 x num_data] for each input vector
and matrix dfce [num_class x num_data] of unnormalized a posterior
probabilities
dfce(y,i) = Conditional_probability(X(:,i)|y)*Prior(y).
If the field model.eps exists then the Bayesian classifier
with the reject option is used. The eps is penalty for the
decision "don't know" which is indicated by label y = 0.
Input:
X [dim x num_data] Vectors to be classified.
model [struct] Describes probabilistic model:
.Pclass [cell 1 x num_classes] Class conditional probabilities.
.Prior [1 x num_classes] A priory probabilities.
.eps [1x1] (optional) Penalty of decision "don't know".
Output:
y [1 x num_data] Labels (1 to num_classes); 0 for "don't know".
dfce [num_classes x num_data] Unnormalized a posterior
probabilities (see above).
Example:
trn = load('riply_trn');
tst = load('riply_tst');
inx1 = find(trn.y==1);
inx2 = find(trn.y==2);
model.Pclass{1} = mlcgmm(trn.X(:,inx1));
model.Pclass{2} = mlcgmm(trn.X(:,inx2));
model.Prior = [length(inx1) length(inx2)]/(length(inx1)+length(inx2));
ypred = bayescls(tst.X,model);
cerror(ypred,tst.y)
See also
BAYESDF, BAYESERR.
About: Statistical Pattern Recognition Toolbox
(C) 1999-2003, Written by Vojtech Franc and Vaclav Hlavac
Czech Technical University Prague
Faculty of Electrical Engineering
Center for Machine Perception
Modifications:
09-jun-2004, VF
01-may-2004, VF
11-mar-2004, VF, "don't" know decision added.
19-sep-2003, VF