"... In PAGE 9: ... Applying the maximum pair consistency probability crite- rion, the three continuous variables are categorized, assuming that qtc is a normal variable, t a log-normal var- iable and bt a variable with a common central tendency. Table4 shows the selected cutoff points and the changes of the pair consistency probability. Comparing the pair consistency probabilities of the categorized variable, we can observe how predictive ability changes with polychot- omization and the pair consistency probability C can be used as a measure to evaluate the loss of predictive ability by categorization.... ..."

"... In PAGE 4: ... The RSA procedure was tested with leave-one-out test- ing (Lachenbruch and Mickey, 1968) to evaluate its ac- curacy in predicting future outcomes. Table1 shows the mean generalization errors of the RSA formulation compared with the following prediction methods: Pooled RSA: All of the feature examples were weighted equally in the objective. See Equation 2.... In PAGE 4: ... In order to simulate feature selection, the k-nearest neighbor algorithm was also tested using only the six features found most relevant by RSA (see pre- vious section). The best results, which are re- ported in Table1 , were obtained using all of the unscaled features, and k equal to seven. Comparative results on all points, recurrent cases only and non-recurrent cases only are shown in Table 1 for the various prediction methods.... In PAGE 5: ...eature selection method described in Section 2.2. Ta- ble 2 summarizes the computational e ect of feature selection, performing the RSA procedure on the Wis- consin prognostic data. As in Table1 , the cumulative error is divided into errors in predicting recurrent and non-recurrent cases. Three di erent styles of feature selection were tested: no selection, tuning-set selection as described above, and a variation which eliminates features down to a certain predetermined number.... ..."

"... In PAGE 4: ... The RSA procedure was tested with leave-one-out test- ing (Lachenbruch and Mickey, 1968) to evaluate its ac- curacy in predicting future outcomes. Table1 shows the mean generalization errors of the RSA formulation compared with the following prediction methods: Pooled RSA: All of the feature examples were weighted equally in the objective. See Equation 2.... In PAGE 4: ... In order to simulate feature selection, the k-nearest neighbor algorithm was also tested using only the six features found most relevant by RSA (see pre- vious section). The best results, which are re- ported in Table1 , were obtained using all of the unscaled features, and k equal to seven. Comparative results on all points, recurrent cases only and non-recurrent cases only are shown in Table 1 for the various prediction methods.... In PAGE 5: ...eature selection method described in Section 2.2. Ta- ble 2 summarizes the computational e ect of feature selection, performing the RSA procedure on the Wis- consin prognostic data. As in Table1 , the cumulative error is divided into errors in predicting recurrent and non-recurrent cases. Three di erent styles of feature selection were tested: no selection, tuning-set selection as described above, and a variation which eliminates features down to a certain predetermined number.... ..."