Where Do Se-trees Perform? (part I)

As a classiier, a Set Enumeration (SE) tree can be viewed as a generalization of decision trees. We empirically characterize domains in which SE-trees are particularly advantageous relative to decision trees. Speciically, we show that: 1. SE-trees excel in domains in which relatively few examples are available; and 2. SE-trees excel in noisy domains. In noisy domains, we discover that SE-trees perform more consistently (measured by the variance in error) in one part of the spectrum, and less consistently in the other; in the lack of noise, we nd that SE-trees are almost invariably more consistent than their decision tree counterparts. Finally, we develop a simple complexity measure based on a target function's syntactic form, and show that SE-trees enjoy a particular advantage in more complex domains.