Table 2 shows the results of feature selection for naive
Bayes; results for naive Bayes with no feature selection
(All features) are shown as well. Accuracies give the
percentage of correct classi¯cations, averaged over the
¯fty trials. Results for CFS are shown in bold if they
show signi¯cant improvement over the corresponding
result for the wrapper, and vice versa. A +" or ¡"
sign shows where results for CFS are signi¯cantly better
or worse than when no feature selection is performed
(all the features are used), and similarly for the wrapper.
Throughout, we speak of results being signi¯-
cantly di®erent" if the di®erence is statistically di®erent
at the 5% level according to a paired two-sided t-test.