Seminar on Computational Learning and Adaptation

  Large Margin Decision Trees

Donghui Wu
Business Intelligence Product Division
Oracle Corp.

Increasing the generalization of decision trees has been a constant topic in machine learning. The recent developments in VC dimension theory and support vector machines reveal that there are connections between generalization error and margin. In particular the overall generalization error of decision tree is bounded by tree size and the margins of separations. In this talk, I will present the concept of margin of separation in decision trees, and introduce several techniques of increasing the margins of decision trees. Initial experimental results show that large-margin version decision trees outperform their regular counterparts. However, the general applicability of these techniques are unclear.

Return to the seminar schedule