DescriptionDecision trees and rules are widely used in different applications as ways for knowledge representation, as parts of algorithms, and as parts of predictors (classifiers). Depending on the problem under consideration we need to minimize or maximize some cost functions: depth, average depth, number of nodes, number of misclassifications for decision trees; length, coverage, number of misclassifications for decision rules. Unfortunately, almost all mentioned optimization problems are NP-hard which means that most probably for these problems there are no polynomial time algorithms.;;To avoid this “restriction” we created in KAUST a software system Dagger which is based on extensions of dynamic programming. This system allows sequential optimization of decision trees and rules relative to different cost functions. It can describe relationships between two cost functions, in particular, between depth and number of misclassifications of decision trees. Experiments show that Dagger is applicable to moderate size decision tables that came from real-life applications.;;The aim of presentation is to acquaint students and researchers in KAUST with possibilities of Dagger which can be used in their study of experimental data. The expected time of the presentation is 90 minutes. The presentation contains simple introduction into used algorithms, and many experimental results.
Mikhail Moshkov is professor in MCSE Division at KAUST since October1, 2008. He earned master’s degree from Nizhni Novgorod State University,received his doctorate from Saratov State University, and habilitation fromMoscow State University.From 1977 to 2004, Dr. Moshkov was with Nizhni Novgorod State Uni-versity. Since 2003 he worked in Poland in the Institute of Computer Science,University of Silesia, and since 2006 also in the Katowice Institute of Infor-mation Technologies.His main areas of research are complexity of algorithms, discrete opti-mization, and machine learning. Dr. Moshkov is author or coauthor of fourresearch monographs published by Springer.
No resources found.
No links found.