Hybrid Decision Tree Learners with Alternative Leaf Classifiers: An Empirical Study

Therehasbeensurprisinglylittle researchso far that systematicallyinvestigatedthe possibilityof constructinghybrid learningalgorithmsbysimplelocal modificationsto decision tree learners. In this paperweanalyzethree variantsof a C4.5-stylelearner, introducingalternativeleaf models(Naive Bayes,IBI, and multi-responselinear regression, respectively) whichcanreplacethe originalC4.5leaf nodesduring reducederror post-pruning.Weempiricallyshowthat these simplemodificationscan improveuponthe performanceof the original decisiontree algorithmandevenuponbothconstituent algorithms.Wesee this as a step towardsthe constructionof learnersthatlocallyoptimizetheir biasfor different regionsof the instancespace.

• Type: Conference
• Year: 2001
• Where: FLAIRS
• Authors: Alexander K. Seewald, Johann Petrak, Gerhard Widmer
• Keywords: Artificial Intelligence, Decision Tree Learners, FLAIRS 2001, OriginalC4.5leaf Nodesduring Reducederror, Step Towardsthe Constructionof