Experiments
Twelve standard datasets drawn from the UCI collection
(Merz and Murphy, 1996) were used in the experiments:
they are summarised in Table 1. These datasets
were chosen because of the prevalence of nominal features
and their predominance in the literature. Three
of the datasets (australian, lymphography, and horsecolic)
contain a few continuous features; the rest contain
only nominal features.
Fifty runs were done for each machine learning algorithm
on each dataset with features selected by CFS
and by the wrapper. In each run, a dataset was randomly
split into a training and testing set (sizes given in
Table 1). CFS and the wrapper were applied in turn to
the full training set to select features. Separate training
and testing sets consisting of features selected by CFS
and features selected by the wrapper were created and
each machine learning algorithm was applied to these
dimensionally reduced datasets.