Optimally Trained Regression Trees and Occam's Razor

0404694 - UIVT-O 20020058 RIV DE eng C - Conference Paper (international conference)
Savický, Petr - Klaschka, Jan
Optimally Trained Regression Trees and Occam's Razor.
COMPSTAT 2002. Proceedings in Computational Statistics. Heidelberg: PhysicaVerlag, 2002 - (Härdle, W.; Rönz, B.), s. 479-484. ISBN 3-7908-1517-9.
[COMPSTAT 2002. Berlin (DE), 24.08.2002-28.08.2002]
R&D Projects: GA ČR GA201/00/1482
Institutional research plan: AV0Z1030915
Keywords : regression trees * recursive partitioning * optimization * dynamic programming * bottom-up algorithms * generalization * Occam's razor
Subject RIV: BA - General Mathematics

Two bottom-up algorithms growing regression trees with the minimum mean squared error on the training data given the number of leaves are described. As demonstraded by the results of experiments with simulated data, the trees resulting from the optimization algorithms may have not only better, but also worse generalization properties than the trees grown by traiditional methods. This phenomenon is discussed from the point of view of the Occam's razor principle.
Permanent Link: http://hdl.handle.net/11104/0124933