The Discovery and Use of Ordinal Information on Attribute Values in Classifier Learning

Rule and tree based classifier learning systems can employ the idea of order on discrete attribute and class values to aid in classifi- cation. Much work has been done on using both orders on class values and monotonic relationships between class and attribute orders. In con- trast to this, we examine the usefulness of order specifically on attribute values, and present and evaluate three new methods for recovering or discovering such orders, showing that under some circumstances they can significantly improve accuracy. In addition we introduce the use of classifier ensembles that use random value orders as a source of variation, and show that this can also lead to significant accuracy gains.