An ID3-like tree-based classifier named RID3 has been proposed. The cl
assifier requires a ranking of the features according to their discrim
inability between classes. We propose a simple but effective scheme fo
r feature ranking. RID3 first constructs a preliminary tree with a def
ault threshold at each node. If the performance of the initial tree is
not satisfactory, then the threshold at each node is tuned with genet
ic algorithms. RID3 is found to outperform nearest-neighbor classifier
for all the data sets considered. (C) Elsevier Science Inc. 1997