Ocenjevanje zanesljivosti napovedi pri ansambelskih metodah

diplomsko delo

Anže Arhar (Author), Zoran Bosnić (Mentor)

Abstract

V diplomski nalogi smo se ukvarjali s področjem izboljšanja točnosti napovedi. Izkoriščali smo pojav spreminjanja variance občutljivostnih napovedi ansambelskih metod ob dodajanju spremenjenih primerov v učno množico. Za vsak testni primer smo izdelali večje število ansamblov, ki pri gradnji modelov uporabljajo naključnost vzorčenja iz učne množice, kot sta metoda bagging in naključni gozdovi. Z dodajanjem spremenjenih testnih primerov smo pridobili občutljivostne napovedi za vsak ansambel. Na podlagi variance občutljivostnih napovedi smo iskali spremembo prvotne napovedi, ki povzroči najboljše ujemanje z ansamblom (varianca napovedi je najmanjša). Preizkusili smo več različnih ansambelskih metod in več metod iskanja minimuma variance napovedi. Eksperimentalno smo dokazali, da pri parametrih, ki smo jih izbrali za evalvacijo, ne obstaja statistična razlika med prvotnimi in popravljenimi napovedmi. S popravljenimi napovedmi nam je uspelo zmanjšati interval zaupanja napovedi za 13 %.

Keywords

napovedovanje;ocenjevanje zanesljivosti;ansambel;regresija;interdisciplinarni študij;univerzitetni študij;diplomske naloge;

Data

Language:	Slovenian
Year of publishing:	2023
Typology:	2.11 - Undergraduate Thesis
Organization:	UL FRI - Faculty of Computer and Information Science
Publisher:	[A. Arhar]
UDC:	004.85(043.2)
COBISS:	158608899
Views:	27
Downloads:	8
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	English
Secondary title:	Assessing the reliability of predictions of ensemble methods
Secondary abstract:	In the diploma thesis, we dealt with the area of improving prediction accuracy. We took advantage of the phenomenon of changing variance of sensitivity predictions when adding modified examples to the training set of ensemble methods. For each test case, we created a larger number of ensembles that use random sampling from the training set, such as random forest and bagging methods to build models. By adding modified test cases we obtained sensitivity predictions for each ensemble. Based on the variance of the sensitivity predictions, we searched for a change to the original prediction that would result in the best match with the ensemble (the variance of the predictions would be the lowest). We tested several different ensemble methods and methods for finding the minimal prediction variance. We experimentally proved that no statistical difference exists between the original and corrected predictions for the parameters we chose for evaluation. We also managed to reduce the confidence interval of the prediction by 13 % with the revised predictions.
Secondary keywords:	prediction;reliability estimation;ensemble;regression;machine learning;computer science;computer and information science;computer science and mathematics;interdisciplinary studies;diploma;Strojno učenje;Računalništvo;Univerzitetna in visokošolska dela;
Type (COBISS):	Bachelor thesis/paper
Study programme:	1000407
Embargo end date (OpenAIRE):	1970-01-01
Thesis comment:	Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages:	42 str.
ID:	19829943