Assessing the Usability of Predictions of Different Regression Models

0348393 - ÚI 2011 RIV SK eng C - Conference Paper (international conference)
Šťastný, J. - Holeňa, Martin
Assessing the Usability of Predictions of Different Regression Models.
Informačné Technológie - Aplikácie a Teória. Seňa: Pont, 2010 - (Pardubská, D.), s. 93-98. ISBN 978-80-970179-3-4.
[ITAT 2010. Conference on Theory and Practice of Information Technologies. Smrekovica (SK), 21.09.2010-25.09.2010]
R&D Projects: GA ČR GA201/08/0802
Institutional research plan: CEZ:AV0Z10300504
Keywords : regression models * confidence of predictions * confidence intervals * transductive inference * sensitivity analysis
Subject RIV: IN - Informatics, Computer Science

Two kinds of methods for assessing reliability of regression models are analysed. The first kind consists in obtaining confidence intervals using either statistical methods (frequentist, Bayesian) or transductive inference. The second kind are heuristic methods, designed so that they correlate with prediction error. Both methods are compared on four kinds of parametric regression models using real-world data. This is a work in progress, that is why transductive inference has not been implemented yet, only one dataset has been used, and the comparison has been implemented for two kinds of regression models, so far: multilayer perceptron and support-vector regression. For them, the variability of results obtained by heuristics methods with respect to different models is illustrated, as well as the correlation between the confidence intervals and the heuristic methods.
Permanent Link: http://hdl.handle.net/11104/0188940