|
NeuroCOLT
Technical Report NC-TR-00-074
Uses
of Convexity in Numerical Domain Partitioning
Tapio Elomaa & Juho Rousu
Abstract
We study the different manifestations of convexity in classifier-constructing
machine learning algorithms. Two principal ways to apply the consequences
of the convexity of the evaluation function to enhance numerical domain
partitioning exist: it enables static and dynamic pruning of partition
candidates. Numerical attribute handling is a potential time-consumption
bottleneck in classifier learning. Therefore, speeding it up is important
for the practical utility of machine learning algorithms. In extensive
empirical evaluation we review the utility of static and dynamic pruning
of partition candidates. We test for 28 UCI test domains the speed-up
gained by both approaches separately and their combined effect. All
pruning methods are able to enhance the efficiency of optimal numerical
attribute partitioning. Best results are obtained by combining static
and dynamic pruning.
Download
Compressed Postscript
|