magistrsko delo
Tomaž Kariž (Author), Janez Demšar (Mentor)

Abstract

V današnjem svetu je zanesljivost napovedovanja zelo pomembna, predvsem na področjih, kot sta recimo zdravstvo in finance, kjer ne bi radi napovedali česa, v kar nismo dovolj prepričani. V strojnem učenju se za reševanje teh problemov raziskuje metode, ki bi nam skušale oceniti, kako zanesljive so naše napovedi. Pri ocenjevanju zanesljivosti napovedi obstajata dve vrsti metod: takšne, ki se specializirajo za točno določen model in takšne, ki ne predpostavljajo vnaprej vrste modela. Prve lahko upoštevajo dodatne informacije pri določanju zanesljivosti, saj lahko uporabijo parametre, ki so specifični za model, kot dodatno informacijo. Druge pa imajo to lastnost, da delujejo na vseh modelih. V delu predstavimo nekaj novih metod, ki delujejo na skupinskih modelih, torej spadajo med tiste, ki so specifične za določen model. Metode delujejo tako na klasifikacijskih kot tudi na regresijskih podatkovnih množicah. Uspešnost metod ovrednotimo s Pearsonovim korelacijskim koeficientom v primeru regresijskih problemov in Wilcoxon-Mann-Whitneyevo statistiko v primeru klasifikacijskih. Razvite metode primerjamo z že obstoječimi in rezultate prikažemo z grafom rangov kritične razdalje.

Keywords

strojno učenje;ocenjevanje zanesljivosti;zanesljivost napovedi;skupinski modeli;računalništvo;računalništvo in informatika;magisteriji;

Data

Language: Slovenian
Year of publishing:
Typology: 2.09 - Master's Thesis
Organization: UL FRI - Faculty of Computer and Information Science
Publisher: [T. Kariž]
UDC: 004.85(043.2)
COBISS: 1536493251 Link will open in a new window
Views: 708
Downloads: 167
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: English
Secondary title: Reliability estimation of ensemble model predictions
Secondary abstract: In today's world, the reliability of a prediction is very important, especially in areas such as health and finance, where we do not want to make predictions that are not sufficiently reliable. To solve these problems in the context of machine learning, methods are being researched that assess the reliability of predictions. There are two types of methods: those specialized for a specific model and those who do not presume in advance the model type. The first may take into account additional information in determining the reliability, because they can use the parameters that are specific to the model as additional information. Others, however, are applicable to all models. In this work, we present some methods that operate on ensemble models, therefore, they are among those that are specific to a particular model. Methods operate on both the classification as well as regression datasets. Performance of methods is evaluated by Pearson correlation coefficient in the case of regression problems and Wilcoxon-Mann-Whitney statistics in the case of classification. The developed methods are compared with existing ones. We also show the results using critical distance diagrams.
Secondary keywords: machine learning;reliability assessment;prediction reliability;ensemble models;computer science;computer and information science;master's degree;
File type: application/pdf
Type (COBISS): Master's thesis/paper
Study programme: 1000471
Embargo end date (OpenAIRE): 1970-01-01
Thesis comment: Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages: 57 str.
ID: 8966322