Show simple item record

Software tool for testing structured regression algorithms based on GCRF model

dc.contributor.advisorDevedžić, Vladan
dc.contributor.otherJeremić, Veljko
dc.contributor.otherStanimirović, Zorica
dc.creatorVujičić, Tijana M.
dc.date.accessioned2018-11-26T14:50:49Z
dc.date.available2018-11-26T14:50:49Z
dc.date.available2020-07-03T09:38:10Z
dc.date.issued2018-07-13
dc.identifier.urihttp://eteze.bg.ac.rs/application/showtheses?thesesId=6161
dc.identifier.urihttps://nardus.mpn.gov.rs/handle/123456789/10194
dc.identifier.urihttps://fedorabg.bg.ac.rs/fedora/get/o:18659/bdef:Content/download
dc.identifier.urihttp://vbs.rs/scripts/cobiss?command=DISPLAY&base=70036&RID=515655066
dc.description.abstractPredmet istraživanja ovog rada su modeli strukturne regresije, koji su dizajnirani da koriste veze između objekata prilikom predviđanja izlaznih vrijednosti. Drugim riječima, modeli strukturne regresije razmatraju atribute objekata i veze između objekata kako bi dali što tačnije predviđanje. Gaussian Conditional Random Fields (GCRF) model je jedan od najčešće korišćenih modela strukturne regresije koji integriše predikciju tradicionalnih modela nadgledanog učenja (nestrukturnih prediktora) i vezu između objekata u cilju tačnije predikcije. Glavna pretpostavka ovog modela je da su dva objekta koja su usko povezana veoma slični jedan drugom i samim tim vrijednosti njihovih izlaznih varijabli treba da budu slični. Sličnost između objekata u GCRF modelu mora da bude simetrična, ali u velikom broju realnih primjera objekti su nesimetrično povezani. U radu je predstavljeno proširenje GCRF modela koje uzima u obzir asimetričnu sličnost između objekata (nazvan usmjereni GCRF - Directed GCRF). Na sintetičkim i realnim setovima podataka pokazano je da novi model daje tačnije predviđanje od standardnog GCRF modela i tradicionalnih nestrukturnih prediktora. Rad obuhvata i razvoj softverskog alata otvorenog koda koji integriše različite vrste GCRF modela i omogućava treniranje i testiranje tih modela na različitim setovima podataka, preko grafičkog korisničkog interfejsa. Izvršena je evaluacija alata sa korisnicima različitih profila i različitog znanja iz oblasti mašinskog učenja. Rezultati su potvrdili da je alat je intuitivan i lak za korišćenje kako za eksperte, tako i za početnike i istraživače iz različitih domena kojima GCRF model može pomoći da dođu do željenih informacija.sr
dc.description.abstractThe subject of this dissertation are structured regression models that are designed to use relationships between objects for predicting output variables. In other words, structured regression models consider the attributes of objects and dependencies between objects to make predictions as accurately as possible. Gaussian Conditional Random Fields (GCRF) model is commonly used structured regression model that incorporates the outputs of traditional supervised learning models (unstructured predictors) and the correlation between output variables in order to achieve a higher prediction accuracy. A main assumption in the GCRF model is that if two objects are closely related, they should be more similar to each other and they should have similar values of the output variable. The similarity considered in GCRF is symmetric. However, in many real-world examples objects are asymmetrically linked. This dissertation presents extension of GCRF model that considers asymmetric similarities between objects (called Directed GCRF). The effectiveness of new model is characterized on synthetic datasets and real-world datasets, on which it was more accurate than the standard GCRF model and baseline unstructured predictors. This dissertation also presents development of an open-source software tool that integrates various GCRF methods and supports training and testing of those methods on different datasets using graphical user interface. The tool was evaluated with users with different level of knowledge in the machine learning field. Evaluation results confirmed that this tool is intuitive and easy to use for experts, as well as for beginners and researchers from different domains that can use GCRF for data prediction.en
dc.formatapplication/pdf
dc.languagesr
dc.publisherУниверзитет у Београду, Факултет организационих наукаsr
dc.rightsopenAccessen
dc.rights.urihttps://creativecommons.org/licenses/by-nc-nd/4.0/
dc.sourceУниверзитет у Београдуsr
dc.subjectinteligentni sistemisr
dc.subjectintelligent systemsen
dc.subjectmachine learningen
dc.subjectstructured regressionen
dc.subjectGCRF modelen
dc.subjectgraphsen
dc.subjectsoftware developmenten
dc.subjectsoftware toolen
dc.subjectopen source softwareen
dc.subjectsoftware usabilityen
dc.subjectmašinsko učenjesr
dc.subjectstrukturna regresijasr
dc.subjectGCRF modelsr
dc.subjectgrafovisr
dc.subjectrazvoj softverasr
dc.subjectsoftverski alatsr
dc.subjectsoftver otvorenog kodasr
dc.subjectupotrebljivost softvera.sr
dc.titleSoftverski alat za ispitivanje algoritama strukturne regresije bazirane na GCRF modelusr
dc.title.alternativeSoftware tool for testing structured regression algorithms based on GCRF modelen
dc.typedoctoralThesisen
dc.rights.licenseBY-NC-ND
dcterms.abstractДеведжић, Владан; Јеремић, Вељко; Станимировић, Зорица; Вујичић, Тијана М.; Софтверски алат за испитивање алгоритама структурне регресије базиране на ГЦРФ моделу; Софтверски алат за испитивање алгоритама структурне регресије базиране на ГЦРФ моделу;
dc.identifier.fulltexthttp://nardus.mpn.gov.rs/bitstream/id/21991/IzvestajKomisije18330.pdf
dc.identifier.fulltexthttps://nardus.mpn.gov.rs/bitstream/id/21991/IzvestajKomisije18330.pdf
dc.identifier.fulltexthttps://nardus.mpn.gov.rs/bitstream/id/21990/Disertacija.pdf
dc.identifier.fulltexthttp://nardus.mpn.gov.rs/bitstream/id/21990/Disertacija.pdf
dc.identifier.rcubhttps://hdl.handle.net/21.15107/rcub_nardus_10194


Files in this item

Thumbnail
Thumbnail

This item appears in the following Collection(s)

Show simple item record