This paper proposes to use neural network ensemble to edit the training data
set for kNN classifiers. In detail, a neural network ensemble is trained from the
original training data set. Then, the class labels of the training examples are
replaced by the labels generated by the neural network ensemble. Experiments
show that such an approach could achieve better editing effect than the Depuration
algorithm does.
This paper also examines the Depuration algorithm and identifies the two
editing schemes it adopted. Through detaching these two schemes, this paper
derives two new editing approaches from Depuration, i.e. RelabelOnly and RemoveOnly.
Experiments show that the editing effect of Depuration is only comparable
to that of RelabelOnly while worse than that of RemoveOnly. This discloses
that the scheme of RemoveOnly does not function in the Depuration algorithm.
Moreover, in some cases simultaneously using the scheme of RelabelOnly and
the scheme of RemoveOnly is even worse than using either of them. Exploring
the reason behind these observations is an interesting issue for future work.