Up: Samplex Table of Contents Previous: Building a classifier Next: Neural network

Statistical Pattern Recognition Classifier

The available statistical pattern recognition classifiers are the linear and quadratic Bayes classifiers. These classifiers are both parametric, using a parametric representation of the classes' distribution.

The quadratic Bayes classifier uses a normal distribution and a second order discriminant function. The linear Bayes classifier uses only a first order discriminant function.

Error evaluations are presented in the console window for the current data sample. The errors are the resubtitution (R error) and leave-one-out (L error) errors for the quadratic and linear Bayes classifiers. A confusion matrix details the errors. The resubstitution error is measured by classifying the database points (used to build the classifier) and counting the number of misclassification. The leave-one-out is computed iterativelly by excluding one point from the data to build the classifier, then testing that point. The Rerror and Lerror are respectivelly the upper-bound and lower-bound error of the classifier.

Figure 9: Error evaluation and confusion matrix for the quadratic Bayes classifier

The confusion matrix shows how the classifier makes error. In Figure 9 for the quadratic classifier, 7 points for class 4 were mistakenly assigned to class 0.

Up: Samplex Table of Contents Previous: Building a classifier Next: Neural network