CFS outperforms the wrapper four times for naive
Bayes and ¯ve times for C4.5, while the wrapper outperforms
CFS three times for both learning algorithms.
Furthermore, as shown by the entries marked with +"
or ¡" in the tables, CFS improves the accuracy of the
learning algorithms more times and degrades accuracy
fewer times than the wrapper does. For naive Bayes,
CFS improves accuracy eight times and degrades accuracy
only once; the wrapper improves accuracy seven
times but degrades accuracy four times. For C4.5,
CFS improves accuracy twice and degrades accuracy
twice; the wrapper improves accuracy three times but
degrades accuracy ¯ve times.
It appears that the wrapper has some di±culty on
datasets with fewer examples. Cross validation accuracy
estimates can exhibit greater variability when
the number of examples is small (Kohavi, 1995), and
the wrapper may be over¯tting these datasets in some
CFS outperforms the wrapper four times for naive
Bayes and ¯ve times for C4.5, while the wrapper outperforms
CFS three times for both learning algorithms.
Furthermore, as shown by the entries marked with +"
or ¡" in the tables, CFS improves the accuracy of the
learning algorithms more times and degrades accuracy
fewer times than the wrapper does. For naive Bayes,
CFS improves accuracy eight times and degrades accuracy
only once; the wrapper improves accuracy seven
times but degrades accuracy four times. For C4.5,
CFS improves accuracy twice and degrades accuracy
twice; the wrapper improves accuracy three times but
degrades accuracy ¯ve times.
It appears that the wrapper has some di±culty on
datasets with fewer examples. Cross validation accuracy
estimates can exhibit greater variability when
the number of examples is small (Kohavi, 1995), and
the wrapper may be over¯tting these datasets in some
การแปล กรุณารอสักครู่..
