Učenje Bayesovih mreža iz cenzuriranih podataka o preživljavanju

doctoral thesis

• Title (english): Learning Bayesian networks from censored survival data
• Author: Ivan Štajduhar
• Mentor: Bojana Dalbelo Bašić (mentor)
• Committee member: Nikola Bogunović (predsjednik povjerenstva)
• Committee member: Dragan Gamberger (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2010-03-26, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing
• Universal decimal classification (UDC): 004 - Computer science and technology. Computing. Data processing

Abstract

Bayesove mreže kao jedan od oblika predstavljanja znanja i nadziranog strojnog učenja, stekle su veliku popularnost zbog učinkovitog zaključivanja i mogućnosti intuitivnog grafičkog predstavljanja uzročno-posljedičnih veza među kovarijatama. Zbog tih karakteristika često se koriste u sustavima potpore odlučivanju u raznim poljima medicine, no njihova je primjena gotovo u potpunosti zanemarena u analizi preživljenja. Razlog tome leži u potrebi za korištenjem posebnih postupaka rukovanja podacima o preživljenju zbog lošeg utjecaja cenzure. Ova se disertacija bavi problemom učenja Bayesovih mreža iz cenzuriranih podataka o preživljenju. Predstavljeno je više poznatih i jedan novi postupak pripremne obrade takvih podataka za potrebe strojnog učenja. Korišteni modeli su dva poznata algoritma za učenje Bayesovih mreža, model naivnog Bayesovog klasifikatora i regresijski model proporcionalnih hazarda. Provedeno je temeljito testiranje simulacijskom studijom i na nekoliko realnih domena iz područja kliničke medicine. Procjena učinkovitosti postupaka pripremne obrade podataka na modelima izvedena je statističkom usporedbom rezultata testova više standardnih metrika strojnog uèenja i metrika analize preživljenja. Pored toga, testirana je i sposobnost ispravnog otkrivanja uzročne strukture algoritama za uèenje Bayesovih mreža uz razlièite postupke pripremne obrade podataka. Rezultati simulacijske studije sugeriraju kako je predložen postupak pripremne obrade podataka odstranjivanjem šuma cenzure, izvrsno rješenje za visokocenzurirane domene.

Abstract (english)

As a form of knowledge representation and supervised machine learning, Bayesian networks have become increasingly popular because of their efficient inference and their inherent ability of intuitive graphical representation of causal interactions among covariates. For these reasons they are often used in decision support systems in various fields of medicine. At the same time they are being almost completely ignored in survival analysis. This is because special data-handling techniques are needed for handling survival data, due to the bad influence of censoring. This thesis deals with the problem of learning Bayesian networks from censored survival data. Several known data-preprocessing techniques for machine learning and a new technique are presented. Models used include two well-known algorithms for learning Bayesian networks, the naive Bayes classifier and the proportional hazards regression model. Thorough testing was conducted on both a simulation study and on several real-world domains from the field of clinical medicine. Assessment of the efficiency of the data-preprocessing techniques on models was conducted using statistical comparison on test results of several standard machine-learning and survival-analysis metrics. The Bayesian network learning algorithms' ability of identifying the correct causal structure after using different data-preprocessing techniques was also tested. Simulation study results suggest that the proposed data-preprocessing technique of reducing censoring noise performs excellently in heavily censored domains.

• Keywords: Bayesova mreža; analiza preživljenja; cenzura; strojno učenje; prognostički model; predstavljanje znanja
• Keywords (english): Bayesian network; survival analysis; censoring; machine learning; prognostic model; knowledge representation
• Language: croatian
• URN:NBN: urn:nbn:hr:168:515563
• Promotion: 2010
• Study programme: Title: Computer Science; Study programme type: university; Study level: postgraduate; Academic / professional title: Doktor znanosti (Doktor znanosti)
• Type of resource: Text
• Extent: i, 166 str. : graf. prikazi
• File origin: Born digital
• Access conditions: Open access
• Created on: 2024-03-19 12:08:17