In [16] we presented a program named PNM which implements
our projection methods (proposed in [15]). The control of PNM is carried
out interactively. After a trial with PNM, the user analyzes the projection
results (error rates, class-histograms, scatter plots) and, directed by
them, chooses the type of projection (parametric or non-parametric) and
the type of constraints for the discriminant vectors (ORTH or FREE) and
defines the values of control parameters for a new trial. The control of
the PNM is easily carried out by the use of instructions. The definitions
and examples of the control instructions are described at length in the
paper. In order to illustrate how the PNM-program can be used we describe
an interactive classifier design concerning differential diagnoses of the
cerebrovascular accident (CVA). In this application, the data set contains
pathologo-anatomically verified CVA-cases with hemorrhages (91 cases) and
infarction due to ischemia (122 cases). The best class-separation was achieved
by the projection obtained by optimization of our non-parametric discriminant
criterion, which was a consequence of the complicated structure of the
CVA data.
|