Realiability estimation in regression supported with meta-learning and principal component analysis

Povzetek

Ključne besede

Podatki

Jezik:	Angleški jezik
Leto izida:	2010
Tipologija:	2.11 - Diplomsko delo
Organizacija:	UL FRI - Fakulteta za računalništvo in informatiko
Založnik:	[M. Radović]
UDK:	004(043.2)
COBISS:	7963988
Št. ogledov:	68
Št. prenosov:	2
Ocena:	0 (0 glasov)
Metapodatki:

Ostali podatki

Sekundarni jezik:	Slovenski jezik
Sekundarni naslov:	Ocenjevanje zanesljivosti regresijskih napovedi, podprto z metaučenjem in metodo poglavitnih komponent
Sekundarni povzetek:	Reliability estimation in regression supported with meta-learning and principal component analysis Reliability estimates for individual predictions can provide important risk-sensitive information. They enable users to distinguish between better and worse predictions which are very important when dealing with decision-critical prediction problems. Unfortunately, when used on different domains and models, reliability estimates perform differently. That is the reason why we require an approach that can foretell which estimate will work best with a specific domain/model pair. In our thesis we have experimented with meta-learning approach to automatically select the best estimate for a specific domain and regression model. We used seven different meta-classifiers on eight regression models and with nine reliability estimates. These reliability estimates represented the class values of the meta-classification process. During the creation of the metadata for meta-learning, we have chosen meta-features that we considered to be most appropriate for this problem. The results showed that the best performing meta-model was random forest. Meta-model neural networks gave the worst results. Additionally, we proposed principal component analysis approach for creation of two new reliability estimates as combinations of existing ones to see if these new estimates would perform. These results were also compared with the results of the existing nine estimates and the best performing meta-classifier. The results showed that the meta-classifier achieved the best results and the estimate BVCK the second best.
Sekundarne ključne besede:	ocene zanesljivosti;zanesljivost;napak napovedi;metaučenje;metoda poglavitnih komponent;napovedi;strojno učenje;računalništvo;univerzitetni študij;diplomske naloge;
Vrsta datoteke:	application/pdf
Vrsta dela (COBISS):	Diplomsko delo
Komentar na gradivo:	Univerza v Ljubljani, Fakulteta za računalništvo in informatiko
Strani:	66 str.
ID:	23960042