Browsing Doctoral Dissertations by Title
-
Maljković, Mirjana (Beograd , 2021)[more][less]
Abstract: Proteins are linear biological polymers composed of amino acids whose structure and function are determined by the number and order of amino acids. The structure of the protein has three levels: primary, secondary and ter- tiary (three-dimensional, 3D) structure. Since the experimental determination of protein 3D structure is expensive and time-consuming, it is important to develop predictors of protein 3D structure properties from the amino acid sequence (pri- mary structure), such as 3D structure of the protein backbone. The 3D structure of the backbone can be described using prototypes of local protein structure, i.e. prototypes of protein fragments with a length of few amino acids. A set of local structure prototypes determines the library of local protein structures, also called the structural alphabet. A structural alphabet is defined as a set of N proto- types of L amino acid length. The subject of this dissertation is the development of models for the prediction of structural alphabet prototypes for a given amino acid sequence using different data mining approaches. As one of the most known, structural alphabet Protein Blocks (PBs) was used in one part of the doctorial re- search. Structural alphabet PBs consists of 16 prototypes that are defined using fragments of 5 consecutive amino acids. The amino acid sequence is combined with the structural properties of a protein that can be determined based on amino acid sequence (occurrence of repeats in the amino acid sequence) and results of predictors of protein structural properties (backbone angles, secondary structures, occurrence of disordered regions, accessible surface area of amino acids) as an input to the prediction model of structural alphabet prototypes. Besides the de- velopment of models for prediction of prototypes of existing structural alphabet, the analysis of the capability of developing new structural alphabets is researched by applying the TwoStep clustering algorithm and construction of models for the prediction of prototypes of new structural alphabets. Several structural alpha- bets, which differ in the length of prototypes and the number of prototypes, have been constructed and analyzed. Fragments of the large number of proteins, whose structure is experimentally determined, were used to construct the new structural alphabets. URI: http://hdl.handle.net/123456789/5236 Files in this item: 1
dthesis.Matf.Mirjana.Maljkovic.pdf ( 47.43Mb ) -
Čugurović, Milan (Beograd , 2025)[more][less]
Abstract: Compilers use program profiles to perform profile-guided optimizations and pro- duce efficient programs. Although dynamic profilers generate high-quality profiles, they have significant drawbacks. They complicate the application build pipeline by requiring two compi- lation steps and an additional profile collection run. Dynamic profilers also consume substantial time and memory and place a heavy burden on developers to create suitable workloads that accurately reflect typical application usage, cover important code paths, and generate well- distributed profiles. In response to the shortcomings of dynamic profilers, modern static profilers employ ma- chine learning (ML) techniques to predict program profiles. However, state-of-the-art ML-based static profilers often rely on handcrafted features that are platform-specific and difficult to adapt across different architectures and programming languages. They also tend to use computation- ally intensive deep neural networks, which increase application compilation time. Moreover, ML-based static profilers can degrade the performance of optimized programs due to inaccurate profile predictions. This dissertation presents GraalSP , an ML-based static profiler that is portable, polyglot, efficient, and robust. GraalSP achieves portability by defining features on a high-level, graph- based intermediate representation and by partially automating the feature extraction process. This design makes GraalSP polyglot, allowing it to predict profiles for programs written in any language that compiles to Java bytecode, such as Java, Scala, or Kotlin. GraalSP is efficient due to its use of a lightweight XGBoost model based on decision trees, and robust because it relies on carefully designed heuristics that correct machine learning predictions and ensure high performance in programs optimized using the predicted profiles. We integrate GraalSP into the Enterprise GraalVM Native Image compiler and evaluate it on 28 benchmarks from the Renaissance, DaCapo, and DaCapo Scala benchmark suites. These suites represent a modern and diverse collection of benchmarks, featuring numerous real-world workloads across a variety of programming paradigms. Our comprehensive evaluation shows that GraalSP achieves a geometric mean speedup of 7.46% in execution time compared to the default compiler configuration, which models program profiles using a uniform distribution. This dissertation also presents a detailed qualitative and quantitative analysis to position and compare the proposed solution against state-of-the-art static profilers. Additionally, to enhance and expand the evaluation and support developers in analyzing GraalSP ’s predictions, this dissertation introduces the GraalSP-PLog tool. This tool allows developers to run the GraalSP static profiler on any program and generate detailed prediction reports, making it easier to inspect individual predictions and identify model mispredictions. Since GraalSP provides substantial performance gains, has minimal impact on binary size and compile time, and includes a modern, fully automated model retraining pipeline, it is well- suited for commercial deployment. As a result, GraalSP has been the default static profiler for the Enterprise GraalVM Native Image compiler since June 2023, consistently improving performance with every build. URI: http://hdl.handle.net/123456789/5778 Files in this item: 1
Milan_Cugurovic_doktorska_disertacija.pdf ( 5.694Mb ) -
Milogradov-Turin, Jelena (Belgrade)[more][less]
URI: http://hdl.handle.net/123456789/126 Files in this item: 1
phdJelenaMilogradovTurin.pdf ( 115.3Mb ) -
Babanić, Mirko (Beograd , 2022)[more][less]
Abstract: The preparatory part of the dissertation, which leads to the basic one, is based on return-variance parameters that represent two key random variables of the model devised by Markowitz. The research used historical data that in themselves reflect all available information absorbed by the financial market, and therefore, we can consider them not only homogeneous but also absolute (for reasons of realization). Therefore, an analytical procedure of approximation by a sixth-degree polynomial was performed on such data, which represent combinations of values of average returns and variances of portfolio returns, thus establishing a relation that is explicitly expressed by an algebraic sixth-degree polynomial equation. After that, further analytical procedure determined the conditions for the existence of both the minimum and the tangent portfolio and redefined the terms: efficient portfolio set, preference toward risk, risk aversion, and indifference line. The central topic of the dissertation, the revision of Tobin 's separation theorem, is formulated and proved through three theorems, one basic and two auxiliary. URI: http://hdl.handle.net/123456789/5593 Files in this item: 1
Mirko Babanic - Disertacija.pdf ( 1.504Mb ) -
Stamatović, Biljana (Beograd , 1999)[more][less]
-
Jelić, Milena (Belgrade)[more][less]
-
Milovanović - Aranđelović, Marina (Beograd , 2016)[more][less]
Abstract: The aim of this dissertation is to present new properties of some classes of contractive mappings, and applications of this results to nonlinear analysis. It contain six chapters. The first chapter gives basic properties of semi-metric spaces. New result is extension of Borel - Lebesque theorem to semi metric spaces. The second chapter contain generalization of Niemytzki’s fixed point theorem. In third chapter the notion of measure of non-compactness are extended to class of semimetric spaces. New fixed point theorem for condensing mappings defined on semi-metric spaces is presented. Further three chapters contains new common fixed points theorems of Mair - Keeler type, new results for accretive mappings and new common fixed point results for mappings defined on probabilistic metric spaces. URI: http://hdl.handle.net/123456789/4482 Files in this item: 1
DisertacijaMilovanovic-Arandjelovic_Marina.pdf ( 1.599Mb ) -
Živković, Miodrag (Belgrade , 1990)[more][less]
-
Ivanović, Lav (Belgrade , 1983)[more][less]
-
Bulatović, Ranislav (Belgrade , 1983)[more][less]
-
Ševarlić, Branislav (Belgrade)[more][less]
URI: http://hdl.handle.net/123456789/141 Files in this item: 1
phdBranislavMSevarlic.pdf ( 5.662Mb ) -
Đuranović-Miličić, Nada (Belgrade , 1980)[more][less]
URI: http://hdl.handle.net/123456789/69 Files in this item: 1
phdNadaDjuranovicMilicic.pdf ( 3.214Mb ) -
Teleki, Đorđe (Belgrade , 1964)[more][less]
-
Šegan, Stevo (Belgrade , 1987)[more][less]
-
Žižović, Mališa (Belgrade , 1980)[more][less]
-
Đurović, Dragutin (Belgrade)[more][less]
-
Ninković, Slobodan (Belgrade)[more][less]
-
Kuzmanoski, Mike (Belgrade)[more][less]
-
Rečkovski, Nikola (Skopje , 1978)[more][less]
-
Graovac, Jelena (Univerzitet u Beogradu , 2014)[more][less]
Abstract: U svetu u kome ˇzivimo, internet i digitalni zapis uˇcinili su da ogromne koliˇcine sirovih podataka postanu dostupne ˇsirokoj javnosti. Jedan ameriˇcki menadˇzer je joˇs davno izjavio: "Raˇcunari su nam obe´cali fontanu mudrosti, a ovo ˇsto smo dobili je poplava podataka" [20]. Sirovi podaci, neadekvatno strukturirani i razliˇcitih formata, sadrˇzaja i kvaliteta su retko od koristi. Neophodno ih je pripremiti, analizirati i na osnovu toga do´ci do informacija i znanja koja na taj naˇcin stiˇcu neprocenjivu vrednost. Istraˇzivanje podataka (eng. data mining) je interdisciplinarno polje infor- matike koje se bavi automatskim ili polu-automatskim otkrivanjem znanja u podacima. Njegov osnovni zadatak je netrivijalna ekstrakcija informa- cija iz podataka, i to informacija koje su implicitne, prethodno nepoznate i potencijalno korisne. Koriste se metode koje su u preseku veˇstaˇcke in- teligencije, maˇsinskog uˇcenja, statistike i sistema baza podataka [97]. Zadaci koji se reˇsavaju u okviru Istraˇzivanja podataka mogu biti prediktivni (klasi- fikacija, regresija, analiza vremenskih serija) ili deskriptivni (klasterovanje, sumarizacija, pravila pridruˇzivanja, analiza redosleda, otkrivanje anomalija). U okviru ove doktorske disertacije bavimo se problemom klasifikacije tek- stova na osnovu njihovog sadrˇzaja. Smatra se da je preko 80% dostupnih informacija saˇcuvano u tekstualnom obliku. Ve´cina informacija je zapisana prirodnim jezikom, odnosno jezikom koji koriste ljudi za svakodnevnu ko- munikaciju. Za oˇcekivati je da ´ce tehnologije automatske obrade podataka zapisanih prirodnim jezikom postati vode´ce u svetu. Glavni doprinos di- sertacije ogleda se u predstavljanju novih metoda za klasifikaciju tekstual- nih dokumenata. Prva metoda predstavlja unapredenje metode razvijene u cilju otkrivanja autorstva teksta [38]. Metoda je zasnovana na predstavlja- nju dokumenta kao profila koji sadrˇzi fiksiran broj n-grama bajtova koji se pojavljuju u dokumentu, i meri razliˇcitosti pomo´cu koje se odreduje klasa kojoj dokument pripada. Ova metoda je jeziˇcki nezavisna i ne zahteva nikakvu prethodnu obradu teksta niti predznanje o sadrˇzaju teksta ili jeziku na kome je tekst napisan. Druga metoda se zasniva na odabranim koncep- tima kao predstavnicima klasa koji se dobijaju iz srpskog wordnet-a, leksiˇcko semantiˇcke mreˇze za srpski jezik. Deo rezultata iz ove disertacije je sadrˇzan u radovima [23, 27, 22, 21, 56, 26, 25, 24] koji su objavljeni, predati za ob- javljivanje ili su u fazi pripreme. Disertacija je organizovana na slede´ci naˇcin. U glavi 1 je prikazan uvod u oblast klasifikacije podataka, u okviru koga su prikazane vrste klasifikacije, procena kvaliteta klasifikacije i primeri primene. Poseban osvrt dat je na klasifikaciju tekstualnih dokumenata. Prikazani su razliˇciti naˇcini predstavljanja dokumenata kao jednog od najvaˇznijiih koraka u procesu klasifikacije. Predoˇceni su i mnogi problemi i izazovi koji se javlja- ju. Prikazani su korpusi klasifikovanih tekstova na srpskom, engleskom, ki- neskom i arapskom jeziku koji ´ce biti koriˇs´ceni u daljem istraˇzivanju. Uvodna glava zavrˇsava se jednim filozofskim pogledom na proces klasifikacije. Glava 2 daje pregled postoje´cih leksiˇckih resursa za srpski jezik [17] koji se razvijaju u okviru Grupe za jeziˇcke tehnologije na Matematiˇckom fakul- tetu Univeziteta u Beogradu. Ideja je da se ukljuˇcivanjem morfoloˇskih, sin- taksiˇckih i semantiˇckih informacija sadrˇzanih u resursima unapredi proces klasifikacije tekstova na srpkom jeziku, kao jednom od morfoloˇski bogatijih jezika. Predstavljeni su korpusi srpskog jezika, elektronski reˇcnik i srpski wordnet kao i raznovrsne tehnologije koje se koriste za njihovu obradu a koje se razvijaju u okviru Grupe. U glavi 3 su prikazane postoje´ce metode maˇsinskog uˇcenja koje su do sada imale veoma uspeˇsnu primenu u procesu klasifikacije. Prikazane su metode zasnovane na drvetima odluˇcivanja, metode zasnovane na pravilima i rastojanju, statistiˇcki zasnovane metode, metode zasnovane na neuronskim mreˇzama i metode zasnovane na podrˇzavaju´cim vektorima. Nove metode za klasifikaciju teksta prikazane su u glavi 4. U okviru prve metode zasnovane na n-gramima bajtova, uvedeni su nova mera razliˇcitosti i novi teˇzinski faktori u odnosu na osnovnu varijantu metode. Teˇzinski faktori su dodeljeni n-gramima u okviru profila klasa, reflektuju´ci znaˇcaj koji n-grami imaju za pripadaju´cu klasu. Smatra se da n-grami koji imaju ve´cu frekvenciju a pripadaju manjem broju klasa imaju ve´ci znaˇcaj za klasu kojoj pripadaju. Uvodenje ovih teˇzinskih faktora rezultovalo je modifikacijom metode na dva naˇcina: modifikacija na nivou mere razliˇcitosti i modifikacija na nivou profila klase. Druga metoda se odnosi na koriˇs´cenje informacija sadrˇzanih u srpskom wordnetu i srpskom elektronskom reˇcniku u cilju klasifikacije teksta na srp- skom jeziku. Ova metoda zasniva se na pridruˇzivanju odabranih koncepata iz srpskog wordnet-a klasama, na osnovu kojih se izraˇcunava mera pripadnosti klasi i vrˇsi pridruˇzivanje dokumenta nekoj od klasa. Rezultati prikazanih novih metoda sumirani su u okviru glave 5. Na srp- skom korpusu je prikazano poredenje prve metode i njenih modifikacija zas- novanih na n-gramima bajtova, karaktera i reˇci. Osnovna varijanta metode i njene modifikacije za n-grame bajtova, testirani su na korpusima na srpskom, engleskom, kineskom i arapskom jeziku, ˇcime je demonstrirana jeziˇcka neza- visnost metode. U okviru Priloga 1 dodatno su predstavljeni svi rezultati dobijeni testiranjem metode za razliˇcite vrednosti parametara, za sve pred- stavljene mere razliˇcitosti, na svim pomenutim korpusima. Druga metoda testirana je samo na korpusu na srpskom jeziku. Poredenje prikazanih rezultata sa drugim rezultataima iz ove oblasti dato je u glavi 6 a glava 7 prikazuje zakljuˇcke i pravce daljeg rada. URI: http://hdl.handle.net/123456789/3746 Files in this item: 1
PhD_JelenaGraovac.pdf ( 3.870Mb )