Öö Blockinøøööþþøøóò Çç Ae Blockinùðø Èöóóððñ× Áò Ðð××׬ Blockin Blockinøøóò

Abstra t Most lassi ation algorithms experien e diÆ ulties when the input-output distribution is irregular (i.e., neighbor examples in the input spa e have di erent values in the output spa e), and sparse. We hara terize the diÆ ulty of a lassi ation problem using a limited number of examples by re-de ning a measure of lass variation to a ount for irregular distributions, and by introdu ing a new measure of example ohesiveness in the input spa e to a ount for (la k of) sparseness. Our empiri al study indi ates these measures orrelate with a ura y performan e.