Decision trees as information source for attribute selection
Kyoko Fukuda, Brent Martin · 2009
Attribute selection (AS) is known to help improve the results of algorithmic learning processes by selecting fewer, but predictive, input attributes. This study introduces a new ranking filter AS method, the tree node selection (TNS) method. The idea of TNS is to determine significant but fewer attributes by searching through the pre-generated decision tree as information source in the manner of a pruning process. To test the performance of TNS, 33 benchmark datasets (UCI) with various numbers of instances, attributes and classes were investigated along with five known AS methods, and the results were tested with the C4.5 (unpruned) and naive Bayes classifiers. The performance, in terms of classification accuracy improvement, reduction in the number of attributes and the size of the generated decision tree are assessed by various statistical analyses for multiple comparisons. TNS was found to provide the most consistent performance for C4.5 and naive Bayes classifiers, and generated unpruned C4.5 trees with selected fewer attributes were generally found to achieve similar quality to pruned C4.5 trees without any attribute selection.