The dissertation discusses the reliability estimation of individual regression predictions. In contrast with average measures for the evaluation of model accuracy, the reliability estimates for individual predictions can provide additional information which could be beneficial for evaluating the usefullness of the prediction (medical diagnosis, financial and control applications). As a novelty, the dissertation proposes a method for reliability estimation of predictions, which is based on the sensitivity analysis approach and is independent of the regression model. New reliability estimates are compared with traditional or adapted reliability estimates. The problem of optimal reliability estimate selection based on the given problem domain and the regression model was also studied using metalearning and internal cross-validation approach. The testing was performed with 8 regression models, with larger number of benchmark problem domains and in a real domain from the area of medical prognostics. The results showed the potential of the proposed methodology in practice.
 

Extended Abstract (PDF)
Dissertation (PDF) - in Slovene
Algorithms in R (ver 1.0, 23 november 2007)
 

Disertacija obravnava področje ocenjevanja zanesljivosti posameznih napovedi regresijskih modelov. V nasprotju z ocenami točnosti regresijskih modelov kot celote nam ocene zanesljivosti posameznih napovedi lahko zagotavljajo dodatno informacijo, ki je lahko v pomoč pri presoji o uporabnosti in posledičnih ukrepih na podlagi teh napovedi (medicinska diagnostika, finančne aplikacije in kontrolni sistemi). V disertaciji je kot novost zasnovana metoda za ocenjevanje zanesljivosti, ki temelji na analizi občutljivosti regresijskih modelov in je neodvisna od regresijskega modela. Uspešnost novih ocen zanesljivosti je primerjana s klasičnimi ali prilagojenimi ocenami zanesljivosti. Obravnavan je tudi problem samodejne izbire ocene zanesljivosti, ki je najbolj ustrezna danemu regresijskemu modelu in problemski domeni. Ta problem obravnavamo z uporabo metaučenja in z notranjim prečnim preverjanjem. Poskusi so bili izvedeni z uporabo 8 regresijskih modelov, na večjem številu problemskih domen in v realni domeni s področja medicinske prognostike. Rezultati testiranj so pokazali potencial uporabe predlagane metodologije v praksi.
 

Razširjeni povzetek (PDF)
Dissertation (PDF) - in Slovene
Algorithms in R (ver 1.0, 23 november 2007)

 © Zoran Bosnić