A Comparison of Robust Model Choice Criteria Within a Metalearning Study

0531580 - ÚI 2021 RIV CH eng C - Conference Paper (international conference)
Vidnerová, Petra - Kalina, Jan - Güney, Y.
A Comparison of Robust Model Choice Criteria Within a Metalearning Study.
Analytical Methods in Statistics. Cham: Springer, 2020 - (Maciak, M.; Pešta, M.; Schindler, M.), s. 125-141. Proceedings in Mathematics & Statistics. ISBN 978-3-030-48813-0.
[AMISTAT 2019: Analytical Methods in Statistics. Liberec (CZ), 16.09.2019-19.09.2019]
R&D Projects: GA ČR(CZ) GA19-05704S; GA ČR(CZ) GA18-23827S
Institutional support: RVO:67985807
Keywords : Automatic method selection * Linear regression * Robust estimation * Robust Akaike criterion
OECD category: Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

The methodology of automatic method selection (metalearning) allows to recommend the most suitable method (e.g. algorithm or statistical estimator) from several alternatives for a given dataset, based on information learned over a training database of datasets. Practitioners have become accustomed to using metalearning in the context of regression modeling, which is useful in a variety of applications in different fields. Still, none of previous metalearning studies on regression targeted at regression complexity issues and the majority of available metalearning studies for regression considered the standard mean square error as the prediction error measure. In this paper, a metalearning study focused on comparing different method selection criteria for the regression task is presented. A prediction rule, recommending the best regression estimator (possibly robust), is constructed over 31 training datasets. These are publicly available datasets, in which the linear model was carefully examined to be suitable. The results with the highest classification accuracy are obtained if the choice of the best estimator is based on robust versions of Akaike information criterion, particularly the version derived from MM-estimators. The work also advocates an implicitly weighted robust prediction mean square error.
Permanent Link: http://hdl.handle.net/11104/0310207