Search

Možnost napovedovanja prihodnjih dogodkov in njihovih posledic je privlačna ideja, a v praksi težko izvedljiva zaradi velikega števila možnih izidov. Ta diploma predstavlja poskus napovedovanja, ki temelji na predpostavki, da se vzorci iz preteklosti ponavljajo. Svetovne dogodke modeliramo kot skupine konceptov, na podlagi katerih jih gručimo v skupine povezanih dogodkov. Iz teh zgradimo podatkovno zbirko za namene napovedovanja, kjer vhodni atributi opisujejo koncepte, ki se pojavijo v posamezni gruči znotraj danega časovnega okna, ciljne oznake pa so koncepti, ki se pojavijo naslednji teden. Na podatkih ocenimo in primerjamo uspešnost različnih modelov za napovedovanje. Naši poskusi pokažejo, da takšno modeliranje povezav med koncepti prispeva koristne informacije za namene napovedovanja prihodnjih dogodkov. Having the ability to predict the course of future events is an attractive idea, but difficult to achieve in practice because of the vast number of possibilities. This diploma presents an attempt at doing so based on the hypothesis that patterns of events from the past tend to repeat. We model real world events as groups of concepts. We cluster events into groups of related events. We then build a dataset from these clusters, where the attributes of each data point represent concepts that occur in a single cluster within a certain time window, and the target labels are concepts that occur the next week. We compare the effectiveness of several different prediction models. Our tests show that relationships between concepts contain useful information for predicting the future course of events.

Analiza podatkov in odkrivanje povezav med entitetnimi tipi znotraj podatkovnih baz sta lahko zelo zahtevni in zamudni opravili. Vsaka baza ima specifično strukturo, raziskovanje katere od uporabnika terja veliko časa. V diplomski nalogi ovrednotimo uporabo rekonstrukcijske napake pri modeliranju z matričnim razcepom kot mero za ocenjevanje stopnje povezanosti med tabelami oz. entitetnimi tipi v podatkovnih bazah. V ta namen smo razvili orodje, ki se poveže na podatkovno bazo in izpiše rangiran seznam relacij med hranjenimi entitetnimi tipi. Uporabnik lahko tako odkrije informativne povezave in usmerjeno raziskuje podatkovno bazo. Data analysis and discovery of relations between connected entity types within databases can be very labour and time intensive. The reason being that every database has its specific structure, which needs to be examined. In this thesis, we evaluated if the reconstruction error of the matrix factorization model can be used to relate tables or entity types within a database. To test this concept, we developed a Python module that connects to a database and returns a ranked list of relations (pairs of entity types). This enables the user to identify the most informative relations and explore them further.

Portal Odprti Podatki Slovenije (OPSI) je vzpostavilo Ministrstvo za javno upravo s ciljem zagotoviti celostni popis podatkovnih zbirk, ki jih vodijo organi javnega sektorja, ter omogočiti objavo zbirk v obliki odprtih podatkov. Vsi podatki na portalu so podatki, evidence in zbirke, ki nastajajo pri delu organov javnega sektorja in so prosto dostopni. V diplomski nalogi smo obdelali dostopne podatke in jih pripravili v obliki za uvoz v program za strojno učenje in vizualizacijo podatkov Orange. Prikazali smo tudi statistiko metapodatkov na portalu OPSI. Vmesnik uspešno pretvori 788 datotek od 19565. Največ virov datotek in sicer 375 izhaja iz področja Vlada in javni sektor. V datotekah s preglednicami je 1038 stolpcev oz. spremenljivk, ki jih delimo v štiri skupine: diskretne (464), nizi (257), zvezne (180) in časovne znamke (137). Večina datotek (13765 datotek), ki jih vmesnik ne more pretvoriti samodejno, je formata html. S pomočjo vmesnika lahko portalu OPSI sporočamo pripombe ali izboljšave glede podatkov in javljamo nepravilnosti na portalu in podatkih. Podatke, ki jih razviti vmesnik samodejno pretvori, lahko uporabimo v programskem orodju Orange in v njih odkrivamo zakonitosti in zanimive vzorce. The portal Odprti Podatki Slovenije (OPSI) was established by the Ministry of Public Administration with the aim to provide an integrated listing of databases managed by the bodies of the public sector and to allow easy publication of data collections in form of open data. All the data provided on the portal are records and collections, which are created by the public sector bodies and are freely accessible. We have preprocessed the freely accessible data and formatted it into the format for Orange data mining, which is a program for machine learning and data visualization. We provide statistics on the available data in the OPSI portal. The interface automatically transforms 788 files out of 19565. The largest source of files, namely 375, is the area of the Government and the public sector. The interface automatically converts data in 1038 columns total, which are grouped into four categories: discrete (464), string (257), continuous (180) and time (137). Most of the files (13765 files) that the interface cannot transform automatically are in the html format. With the interface we can provide comments or improvements regarding the published data and report any irregularities of the portal and data. The data provided by the interface can be mined for interesting patterns using the Orange data mining software.

Molekularni biologi in genetiki se dandanes ukvarjajo s proučevanjem zapisa DNA. Pri eksperimentih tipično dobijo podatke, ki se zaradi svoje narave težko interpretirajo. Klasične metode primerjajo eksperimentalno dobljeno skupino genov z znanimi, funkcijsko povezanimi skupinami genov, a pogosto niso uspešne. Izboljšava pristopa je metoda SANTA, ki vrednoti eksperimentalne rezultate na podlagi omrežja genov. Metodo SANTA smo implementirali v sklopu diplome, v dodatku Orange Bioinformatics. Orange je že dobro uveljavljen program za obdelavo podatkov. Dodatek Bioinformatics zavzema specializirane funkcionalnosti za obdelavo genskih podatkov. Dodatek smo obogatili tudi z izboljšanim dostopom do zbirk BioGRID in STRING, ki hranita podatke o eksperimentalno določenih povezavah med geni. Molecular biologists and geneticists are, nowadays, mostly focused on studying and understanding the DNA transcription. The gathered experimental data is difficult to interpret. Classical methods compare the discovered groups of genes with predefined gene sets. Unfortunately, these methods do not perform well. A solution is proposed by the SANTA method, which evaluates the experimental results on discovered gene sets based on known and extensive gene networks. In this thesis, we have implemented the method in Orange Bioinformatics. Orange is a well-known data-analysis programme. The Bioinformatics add-on includes specialised functionality for processing genomic data. Furthermore, we have also enriched its widgets for access to BioGRID and STRING, which store information on functional sets of genes and interactions.

V diplomskem delu opišemo sistem za obdelavo in klasifikacijo zvočnih posnetkov, ki smo ga zasnovali in razvili kot dodatek za program Orange. Orange je odprtokodni sistem, namenjen podatkovnemu rudarjenju. Vsebuje gradnike za prikazovanje in manipuliranje s podatki, gradnjo napovednih modelov, in gradnike za mnoga druga opravila. Ne vsebuje gradnikov za branje in obdelovanje zvočnih zapisov. Zato smo v ta namen izdelali dodatek, imenovan ``Audio - IJS,'' znotraj katerega smo razvili osem gradnikov, ki omogočajo manipulacijo z zvokom in integracijo z ostalimi gradniki v Orange. Izdelani gradniki omogočajo filtriranje, segmentacijo (razrez), spektralno analizo, ekstrakcijo značilk, predvajanje zvoka in še nekaj ostalih funkcij, ki so potrebne za povezovanje s preostalimi gradniki Orange. Razviti dodatek uporabniku omogoča hitro izdelavo sofisticiranih analiz zvočnih posnetkov. In this diploma thesis we will describe audio clip analysis and classification system that we designed and developed as an add-on for Orange software. Orange is an open-source system, designed for data mining. It contains widgets for displaying and manipulating data, building predictive models and several other tasks. It does not, however, contain widgets for reading and processing audio clips. For this purpose, we built an add-on called »Audio-IJS« within which we developed eight widgets that enable audio manipulation as well as integration with other Orange widgets. The constructed widgets enable filtration, segmentation (cutting), spectral analysis, feature extraction, audio play as well as several other functions, necessary for connecting with other Orange widgets. The developed add-on enables the user a rapid creation of sophisticated analyses of audio clips.

Število brezžičnih naprav danes hitro narašča, posledično se viša stopnja nasičenosti radijskega spektra, saj soobstoj številnih tehnologij povzroča motnje v omrežjih. Zmogljivost obstoječih tehnologij lahko povečamo z razvojem natančnejših cenilk kakovosti povezav. V diplomskem delu predlagamo, uvedemo in ovrednotimo nov pristop za razvoj sistema za napovedovanje kakovosti povezav, ki temelji na gradnji in izpeljavi značilk. Po predhodni analizi množic podatkov o paketih Wi-Fi in Sigfox tvorimo nove značilke in izgradimo klasifikacijski model za napovedovanje kakovosti povezave. Predlagani modeli se razlikujejo glede na točnost in popolnost napovedovanja posameznih vrst povezav, predvsem povezav srednje kvalitete. Najboljši model pravilno uvrsti 95% testnih primerov, kar je bistveno izboljšanje v primerjavi s 60% točnostjo večinskega klasifikatorja. The number of wireless devices is increasing rapidly. The wireless spectrum is thus becoming crowded as various technologies co-exist and interfere with each other. One possible way to improve the performance of existing technologies is to develop accurate link quality estimators. In this thesis, we propose, implement and evaluate a novel approach to link quality prediction based on feature engineering. Following a preliminary analysis of a dataset with Wi-Fi packet traces and a dataset with Sigfox packet traces, we developed new features and built a classification model for link quality prediction. The proposed models vary in performance with respect to accuracy and completeness of predicting different types of links, mainly links of intermediate quality. The best proposed model achieved 95% classification accuracy, which is a substantial improvement compared to the 60% accuracy of the majority classifier.

Biti lastnik plovila ali flote plovil je stresna naloga, še posebej, če s plovilom upravlja nekdo drug ali pa je plovilo nenadzorovano privezano v marini. Zgodi se lahko marsikaj. Plovilo lahko trči v drugo plovilo, obalo ali drug objekt. Za oddaljen nadzor plovil obstaja mnogo naprav, ki so zmožne lastniku posredovati marsikatero informacijo. Vendar taka, ki bi omogočala zaznavo trka, še ni ponujena na tržišču. V ta namen se v zadnjem obdobju v naprave vgrajujejo pospeškometri. Zgolj podatki, ki jih ti senzorji beležijo, niso prav koristni, dokler se jih ne osmisli. Zato je cilj diplomske naloge osmisliti pridobljene podatke. Ker trki na morju niso prav pogosti, je večina podatkov, ki se lahko kvalificirajo kot trk, lažno pozitivnih. Za boljše razumevanje smo podatke iz senzorjev povezali s podatki o vremenu, poiskali korelacije med meritvami pospeškometra in ostalimi atributi ter klasificirali vsako večje odstopanje vrednosti pospeškometra. Uspeli smo razložiti približno tri četrtine takih točk in tako pripraviti podatke na nadaljnjo obdelavo, najverjetneje z metodami strojnega učenja. Owning a vessel or a fleet can be stressful, especially if someone else is steering it or it is left unsupervised in the marina. Many things can go wrong and a collision with another vessel, the shore or other objects is possible. Of course there are many devices, created especially to monitor one's vessel that are capable of transmitting numerous different informations to the owner, but none of those are capable to detect collisions. Lately, accelerometers are being implemented into these devices, but the data they provide is not useful until it is processed properly. The goal of this thesis is to process vessel sensor data and classify it. Because collisions at sea are fairly rare, most of the data recognised as a collision is false positive. To better understand what exactly is going on with the vessel at any given moment, the accelerometer and all other vessel data is merged with weather data. After that, correlation between accelerometer data and any other given attribute is calculated. Outliers are then classified, according to the previous analysis. Roughly three quarters of all outliers were classified correctly and the data is prepared for further processing, most probably by applying machine learning algorithms.

Cilj diplomskega dela je izgradnja modela, ki napoveduje trende vrednosti kriptovalut na podlagi podatkov spletne borze Poloniex. Problem je zanimiv zaradi potencialnih možnosti avtomatskega trgovanja s kriptovalutami, ki bi jih uspešen model omogočal. Za napovedovanje smo uporabili novice iz obravnavanega področja, ki so bile pridobljene iz spletne strani Reddit. Poleg tekstovnih podatkov so za napovedovanje uporabljeni tudi numerični podatki vrednosti kriptovalut pred objavo novice. Problem smo zastavili kot trirazredno klasifikacijo. Z uporabo metode TF-IDF, informacije o sentimentu, polariteti ter podatkih o člankih in stanju trga pred objavo besedila je bila dosežena 50,8% klasifikacijska točnost na validacijski množici ter 49,3% klasifikacijska točnost na testni množici. Za učenje in napovedovanje smo uporabili metodo podpornih vektorjev. Kljub relativno dobri klasifikacijski točnosti, model v praksi verjetno ne bi dajal dobičkonosnih napovedi. The aim of this work is to build a model that predicts cryptocurrency trends based on data from the Poloniex online market. The problem is interesting because of the potential for automated cryptocurrency trading that a successful model would allow. For the forecast, we used the news from the subject area, which were obtained through the Reddit website. In addition to textual data, numerical market data before publishing the news is also used for forecasting. We approached the problem as a three-class classification problem. Using the TF-IDF method, sentiment information, polarity, and article and market information before the publication of the text, we achieved 50,8% classifying accuracy on the validation set and 49,3% classification accuracy on the test set. We used support vector machines for learning and prediction. We have found that in practice, despite the significant classification accuracy, the model is unlikely to yield profitable returns.

Cilj diplomskega dela je bil izdelati spletno aplikacijo, ki deluje kot grafični vmesnik za uporabnike microCOMB-a in vzdržuje bazo genskih ekspresij. Glavne funkcije aplikacije so omogočiti uporabnikom posredovanje ekspresijskih podatkov v analizo in prikazati njene rezultate, vodenje zgodovine analiz in skrbeti za ažurnost javne baze ekspresijskih podatkov. V delu so opisane uporabljene tehnologije, arhitektura sistema, razvojni proces ter končna funkcionalnost aplikacije. Ob razvoju smo strmeli k čim bolj modularni arhitekturi ter preprostosti nadaljnjega nadgrajevanja in vzdrževanja. Uporaba zabojnikov Docker nam je omogočila visoko stopnjo modularnosti in neodvisnosti od programske in strojne opreme. Aplikacija je sestavljena iz strežniškega dela in uporabniškega vmesnika. Strežniški del je razvit z uporabo programskega jezika Python z ogrodjema Pyramid in Cornice, uporabniški vmesnik pa z ECMAScript6 in ogrodjema React in Redux. Za hranjenje podatkov na strežniku smo se odločili za kombinacijo relacijske baze PostgreSQL in tekstovnih datotek. Za dodatno stopnjo modularnosti smo na strežniku uporabili SQLAlchemy, kar aplikaciji omogoča, da je agnostična do uporabljene tehnologije RDBMS. The goal of this thesis is to develop a web application that functions as user interface for microCOMB and manages it's gene expression database. The main functions of the application are to enable the user to upload expression profiles to be analyzed and show it's result, store user history of completed analyses and keep the public database up to date. In the thesis we describe the technologies used, architecture, development process and application functionality. During the development and design process we focused on modularity, maintainability and extendability. Using Docker containers we achieved a high degree of modularity and decoupling from the underlying hardware and software. The application is split into server and client side. The server side is developed using Python and two frameworks Pyramid and Cornice. Client side uses ECMAScript6 as the main language and React in Redux frameworks. To store data on the server side we use a combination of PostgreSQL and text files. To add another degree of modularity we used SQLAlchemy as the ORM on the server side. Using an ORM we made the application RDBMS agnostic.

Interakcije med proteini in RNA imajo pomembno vlogo pri uravnavanju genske ekspresije in posledično na delovanje celic. Napake v interakcijah so pogosto povezane z nastankom bolezni, kot so nevropatije, rak, itd. Poznavanje mest interakcij je tako nujno za razumevanje, odkrivanje, uravnavanje genske ekspresije in zdravljenje omenjenih bolezni. V magistrskem delu smo se osredotočili na modeliranje mesta interakcije proteinov z RNA na osnovi simuliranih podatkov metode RBDmap, ki je nadaljevanje študije Castella in sodelavcev, objavljene leta 2012. Podatke RBDmap smo simulirani na podlagi zbirke PDB, ki hrani strukture 3D kompleksov proteinov in RNA. Za napovedovanje posameznih aminokislin oziroma krajših zaporedij v fragmentih smo preizkusili vrsto metod strojnega učenja, kot so metoda podpornih vektorjev, klasifikacijska drevesa, naivni Bayesov klasifikator in K-najbližjih sosedov. Razvili smo tudi metodo, ki določi aminokisline v interakciji z RNA na podlagi lastnosti fragmentov aminokislin in celotnega proteina. Uspešnost metode je primerljiva s trenutno obstoječimi metodami (AUC 0,783). V nasprotju s pričakovanji, opisovanje fragmentov v splošnem ni pripomoglo k izboljšanju napovednih modelov. Interactions between proteins and RNA play an important role in the regulation of gene expression and therefore in the functioning of cells. Errors in interactions are often related to the development of diseases, such as neuropathy, cancer, etc. To this end, knowing the locations of interactions is crucial for understanding, discovering and managing gene expression and for treating those diseases. The master's thesis focuses on modeling the amino acids interacting with RNA based on simulated data on RBDmap experiments, which is the continuation of the study by Castello et al. from 2012. RBDmap was simulated using the PDB database on 3D structures of ribonucleoprotein complexes. A number of methods of machine learning, such as support vector machines, classification tree, naive Bayes classifier and k-nearest neighbours were evaluated for predicting individual amino acids and fragments of amino acids interacting with RNA. Moreover, a method was developed to determine amino acids interacting with RNA, which considers the characteristics of fragments of amino acids and the entire protein. The method achieved good results (AUC 0.783), which is comparable with current methods. Including features on fragments did not improve the predictive model.

V diplomskem delu smo uporabili različne metode strojnega učenja za uvrščanje virusnih zaporedij v ustrezne taksonomske skupine. Z dostopanjem do podatkovne zbirke NCBI, ki hrani biološke in biotehnološke podatke, smo najprej sestavili celotno taksonomsko strukturo znanih virusnih zaporedij. Podatke smo ustrezno filtrirali in tako zgradili množico učnih primerov. Nato smo uporabili klasične metode strojnega učenja in metodo strukturiranega napovedovanja in ovrednotili uspešnost napovedovanja v taksonomske skupine. V delu smo preučili, kateri načini opisovanja genomskih zaporedij so najprimernejši. Opis genomskih zaporedij s k-terkami ne zajame vseh podrobnosti genomov, zato so najboljši doseženi rezultati le nekoliko boljši od večinskega klasifikatorja. Predznanje o evolucijski povezanosti taksonomskih skupin nekoliko izboljša napovedi modelov, ki to znanje lahko uporabijo. In this diploma thesis our goal was to classify viral sequences into taxonomic groups by using different machine learning methods. We assembled the taxonomic structure by collecting data from NCBI web site. To clean the data we applied several filtering steps. We then evaluated the predictive performance of classical and structured machine learning methods on the task of classification in taxonomy groups. We wanted to determine the most suitable way to describe genomic sequences. Using k-mers to describe the genomic composition yielded poor predictive models, with best performance slightly above the performance of the majority classifier. Methods, which are able to use prior knowledge on the taxonomic relationships between classes, performed slightly better than methods, which did not use such information.

Atributi, s katerimi opisujemo primere v bazah podatkov, so pogosto zelo številni. Določanje resnično pomembnih atributov za klasifikacijo ter njihovih medsebojnih odvisnosti zato predstavlja velik izziv. Eden od načinov, kako zmanjšati dimenzionalnost prostora in določiti pomembne atribute in primere, je z uporabo nenegativne matrične faktorizacije. V magistrski nalogi smo najprej preučili osnove nenegativne matrične faktorizacije in nekaj načinov prikaza podatkov in faktorskih modelov v matrikah. Predlagamo nekaj načinov, kako prikazati in razumeti modele, pridobljene s faktorizacijo. Uspeh metod smo ovrednotili na nekaj podatkovnih zbirkah in ugotovili, da nam vsaka metoda razkrije uporabne informacije o modelu. Z gručenjem faktoriziranih matrik lahko dobimo čistejše gruče kot z gručenjem izvornih podatkov. S projekcijo primerov v prostor faktorjev lahko ugotovimo, kateri faktorji vplivajo na določene razrede. Če pa tej projekciji dodamo še atribute, lahko sklepamo še o povezavi med primeri in atributi izvornega prostora. Attributes that describe data in the databases present themselves in large numbers. For this reason defining truly important attributes for classification and establishing their mutual dependence poses a significant challenge. One way of reducing the dimensionality of the space and defining important attributes and examples is by using non-negative matrix factorization. In this master thesis we first examined the basics of non-negative matrix factorization and a few ways of visualizing the data and factor models in matrices. We propose a few ways of presenting and understanding the models acquired with factorization. We evaluated the effectiveness of the methods on several databases and learnt that each method reveals useful information about a model. Clustering of the factorized matrices can produce purer clusters than clustering of the source data. By projecting examples to the factor space we can see which factors affect certain classes. Adding attributes to this projection makes it possible to deduce the link between the examples and the attributes of the source space.

Cilj diplomske naloge je bil implementirati sekvenčni algoritem za iskanje krajših zaporedij v genomih, ga pohitriti s paralelizacijo ter implementirati paralelno verzijo na grafični kartici. Sekvenčni algoritem je moral poiskati krajša zaporedja znakov v danem zaporedju, ki predstavlja genom. Izračunati je moral frekvence pojavitev zaporedij in pogostost interakcij na podanih položajih ter na naključno premaknjenih položajih, za vsako krajše zaporedje posebej. Na podlagi pojavitev je nato moral določiti, katera krajša zaporedja so za določene položaje v genomu bolj značilna. Na podlagi podatkov o interakcijah med proteini in genomom na določenih položajih, ter na podlagi najdenih krajših zaporedij znakov, je moral nato izračunati in statistično ovrednotiti pogostost pojavitve interakcij. Sekvenčni algoritem smo implementirali v programskem jeziku C, paralelizacijo sekvenčnega algoritma pa smo izvedli na podlagi arhitekture OpenCL, ki omogoča implementacijo algoritmov na grafičnih karticah. The goal of this thesis was to implement a sequential algorithm that would search for subsequences in a genome. To accelerate the execution time of this algorithm we designed a parallel version and implemented the parallel version on a graphics card. The sequential algorithm had to search for predefined subsequences in a genome that was represented as a sequence of characters. It had to calculate the frequencies of sequence occurrences and the frequencies of interactions on predefined positions and on randomly modified positions in the genome, for each subsequence. Based on these frequencies it had to identify sequences that were more frequent on certain locations in a given genome. Based on data about protein-RNA interactions on certain locations in the genome, and based on the found character sequences, the algorithm had to calculate and statistically evaluate the frequencies of interactions. The sequential algorithm was implemented in the C programming language, while the parallelization was implemented on the OpenCL architecture.

Matrično faktorizacijo, ki se povezuje s postopkom zlivanja podatkov, uporabljamo za odkrivanje vzorcev oziroma skupin v podatkih. Faktorizirani model preslika podatke v nižje-dimenzionalen prostor, jih tako skrči in odpravi del šuma. Tovrstni modeli so zato navadno bolj robustni in imajo višjo napovedno točnost. Pri nevronskih mrežah bi tako znali reševati problem prevelike prilagojenosti podatkom (angl. overfitting) in pridobili pri generalizaciji. V nalogi smo preučili, ali s hkratno faktorizacijo parametrov nevronske mreže, ki jih je možno predstaviti z več matrikami, odstranimo (porežemo) nepomembne povezave in tako izboljšamo napovedno točnost mreže. Predlagani postopek rezanja smo preizkusili na navadnih in globokih nevronskih mrežah. Po uspešnosti je primerljiv z ostalimi najuspešnejšimi standardnimi pristopi rezanja nevronskih mrež. Matrix factorization and the procedure of data fusion are used to detect patterns in data. The factorized model maps the data to a low-dimensional space, therefore shrinking it and partially eliminating noise. Factorized models are thus more robust and have a higher predictive accuracy. With this procedure we could solve the problem of overfitting in neural networks and improve their ability to generalize. Here, we report on how to simultaneously factorize the parameters of a neural network, which can be represented with multiple matrices, to prune not important connections and therefore improve predictive accuracy. We report on empirical results of pruning normal and deep neural networks. The proposed method performs similarly to the best standard approaches to pruning neural networks.

V magistrskem delu smo razvili modele za napovedovanje mest na RNA v interakciji s proteini, pri čemer smo uporabili metodo skritih Markovih modelov. Določili smo reprezentativne atribute in motive, na podlagi katerih smo nato zgradili modele za vsak posamezen eksperiment. Ugotovili smo, da različni proteini uporabljajo veliko skupnih atributov in motivov za prepoznavo mest interakcije z RNA. Modeli z dvema stanjema (prisotnost interakcije) boljše napovedujejo kot pa modeli s tremi stanji (prisotnost in intenziteta interakcije). Združevanje napovedi modelov posameznih eksperimentov ne izboljša zmogljivosti napovedovanja. Združeni modeli imajo dovolj dobro zmogljivost, da nam lahko služijo za sklepanje o relacijah med posameznimi proteini, in sicer njihovem sodelovanju, tekmovanju ali neodvisnosti. Skriti Markovi modeli predstavljajo primerno metodo za napovedovanje mest interakcij. The main goal of the master's thesis has been to develop prediction models for interactions between RNA and proteins. We have chosen hidden Markov models as our method for modelling and predicting interactions. From our initial data we have extracted representative features and motifs, which we used for building separate models for each experiment. Majority of proteins bind to the same or very similar features and motifs. We have compared the predictive accuracy of models build with two (presence of interaction) and three states (presence and intensity of interaction). Results show that models with two states perform better than models with three states. Merging predictions of multiple single experiment models in combined models, does not improve prediction accuracy. However, combined models perform with high accuracy, and can be used to determine the relations between proteins, such as competition, cooperation and independence with other proteins when interacting with RNA. We have presented hidden Markov models as viable method for predicting interactions between RNA and proteins.

V diplomski nalogi smo modelirali povezavo med genotipom in fenotipom tridesetih vzorcev kvasovke S. cerevisiae. Na podlagi podatkov in predznanja smo določili mutacije posameznih nukleotidov in z njimi povezane gene, s katerimi je možno zgraditi dober model za napovedovanje fenotipa. Poleg določanja pomembnih mest v genomu (SNV-jev) nam zgrajeni model omogoča tudi določevanje pomembnih genotipov oziroma starševskega izvora, ki je povezan z opazovanim fenotipom. Vrednotenje modelov pokaže, da lahko z linearno regresijo zanesljivo napovedujemo fenotip. Fenotip relativno dobro napoveduje tudi model, ki je zgrajen le na podlagi podatkov o dveh izvornih starših in začetne populacije. Empirično smo določili povezavo med številom vzorcev, ki jih uporabimo za izgradnjo napovednih modelov, in napovedno napako modelov. We have modeled the relationship between genotype and phenotype using data on thirty yeast S. cerevisiae samples. Using prior knowledge, we have determined mutations of individual nucleotides and related genes with which it is possible to build a good prediction model for the phenotype. The constructed models allow us to determine the location of important mutations in the genome (SNVs), to rank samples based on phenotype, and to determine signi_cant genotypes or parental origin, which is connected to the observed phenotype. Evaluation of these models shows that the phenotype can be predicted very reliably with linear regression. The phenotype can be predicted relatively well from data on two starting parents and the _rst pool of segregants. We also show the relation between the number of samples used to build a predictive model and its predictive error.

Diplomsko delo obravnava aplikacijo za prikaz in analizo podatkov o kakovosti električne energije v elektroenergetskem omrežju. Delo je razdeljeno na tri dele. V uvodnem delu opišemo elektroenergetska distribucijska omrežja in definiramo pojem kakovosti električne energije. Nato opišemo splošne značilnosti sitemov za nadziranje kakovosti električne energije (angl., power quality monitoring system, PQMS), tipe podatke in merilne naprave, ki omogočajo merjenje kakovosti električne energije v distribucijskem elektroenergetskem sistemu. V zadnjem delu opišemo razvito spletno platformo MiSmart za prikaz in analizo podatkov o kakovosti električne energije, kar je tudi glavni prispevek diplomskega dela. Opišemo del aplikacije imenovan „Pregledovalnik podatkov," ki smo ga v celoti implementirali v okviru diplomske naloge. Platforma MiSmart omogoča hitrejši odziv na težave v elektroenergetskem omrežju, učinkovitejše vzdrževanje in lažje zagotavljanje kakovosti elektroenergetskega omrežja. In this thesis we describe an application for collecting and analyzing power quality data of electric distribution systems. The thesis is divided into three parts. We begin with a theoretical description of electric distribution networks and define power quality. We then describe the basics of quality monitoring systems and measuring devices, which are used to measure power quality in electric distribution network. The main contribution of this thesis is an implementation of MiSmart, a web platform for monitoring and analysis of power quality data. We describe the part of application named „Data monitor", which we developed in this thesis. MiSmart platform allows faster response to problems, efficient maintenance and provides better power quality in electrical distribution networks.

Interakcije med proteini in RNA imajo ključno vlogo pri velikem številu celičnih procesov. Eksperimentalna analiza 3D struktur molekul je počasna in zahtevna, zato obstaja velika potreba po računskih metodah, ki uspešno napovedujejo mesta ter strukturo molekul v interakciji. V magistrskem delu smo definirali vrsto značilk, ki opisujejo lokalne lastnosti interakcij protein-RNA, na podlagi podatkov o 3D strukturah molekul protein-RNA. Razvili smo metodo, ki združuje strojno učenje in optimizacijski postopek za napovedovanje mesta interakcij med proteinom in RNA. Napovedi strojnega učenja se uporabijo za določanje začetnega stanja optimizacije. Optimizacijski postopek nato uporabi ocenjevalne funkcije osnovane na porazdelitvi 3D strukturnih značilk in tako predlaga najverjetnejšo pozicijo molekule RNA. Predlagani napovedni model dosega natančnost, ki je primerljiva z uspešnostjo najboljših obstoječih metod. Protein-RNA interactions have an essential role in many cellular processes. Experimental analysis of 3D molecular structure is slow and difficult process. Consequently, computational methods, which successfully predict interaction sites and molecular conformations are needed. In this thesis we have defined a number of attributes to describe local properties of protein-RNA interactions using data on 3D structure of protein-RNA molecules. We have implemented a method that uses machine learning and optimization algorithm for prediction of protein-RNA interaction sites. Machine learning predictions are used to generate initial positions for optimization. Optimization algorithm uses scoring functions based on the distribution of 3D structural attributes to identify most likely positions of the RNA molecule interacting with a given protein. The accuracy of the proposed prediction model is comparable to results obtained with best existing methods.

Proteini so pomembni akterji v mnogih celičnih procesih. Interakcije proteinov in RNA imajo pomembno vlogo pri uravnavanju izražanja genov in posledično njihove funkcije. V interakcijo z RNA navadno vstopa več tako posameznih proteinov kakor tudi skupin proteinov. Metoda iCLIP omogoča do nukleotida natančno detekcijo mest interakcij na RNA z izbranim proteinom. V diplomskem delu smo razvili metodo, ki sprejme nabor mest na RNA, ki so v interakciji z izbranim proteinskim kompleksom. V okolici teh mest nato poišče vzorce interakcij iz nabora interakcij proteinov z RNA, za katere imamo podatke. Najdeni vzorci vključujejo tako posamezne komponente (proteine) danega kompleksa kakor tudi proteine, ki danega kompleksa ne tvorijo, ampak vseeno vplivajo na interakcije kompleksa z RNA. Razvita metoda temelji na postopeku faktorizacije nenegativnih matrik. Metodo smo uspešno preizkusili na podatkih o kompleksu SmB, kjer smo uspešno določiti nekaj proteinov, ki so soudeleženi v kompleksu SmB. Proteins play an important role in many processes in a cell. Protein-RNA interactions greatly affect the balancing of gene expressions and consequently their functions. Interaction with RNA may occur with a single protein or a protein complex. iCLIP method is able to detect the protein-RNA crosslink spots with nucleotide resolution. We developed a method that takes a set of crosslink spots that intersect with the choosen protein complex as an input. Method is searching the crosslink spots neighbourhood for the protein-RNA interaction patterns. Found patterns include parts (proteins) of the choosen complex and other proteins that are not a part of the choosen complex, but still affect the interactions of the choosen complex with the RNA. Method that we developed is based on the non-negative matrix factorization. We successfully tested the method on a SmB complex where we found a few proteins that cooporate with SmB complex.

Filogenetsko drevo prikazuje evolucijska razmerja med taksonomskimi enotami. Danes jih gradimo programsko z uporabo različnih metod in na podlagi različnih virov podatkov (npr., zaporedja DNA ali RNA). To mnogokrat privede do nasprotujočih si evolucijskih hipotez, ki jih nato skušamo združiti v eno samo hipotezo oziroma konsenzno drevo. V diplomskem delu je obravnavana konsenzna metoda asimetričnega srednjega drevesa. Le-ta in še dve aproksimacijski metodi so bile implementirane v odprtokodnem paketu Biopython. Uspešnost metode asimetričnega srednjega drevesa je bila primerjana z uspešnostjo metod za gradnjo striktnega, večinskega in Adamsovega konsenza, na nekaj umetnih in enem dejanskem primeru. Pridobljena drevesa so bila ovrednotena z uporabo Robinson-Fouldsove mere in mere razrešenosti drevesa. Asimetrično srednje drevo je velikokrat najmanj podobno vhodnim drevesom, a hkrati najbolj razrešeno drevo, saj podaja največ informacije o evolucijskih razmerjih med taksonomskimi enotami. A phylogenetic tree displays evolutionary relationships between taxonomical units. Nowadays, they are constructed programmatically by using different methods of inference and relying on various data sources, such as DNA and RNA sequences. The resulting evolutionary hypotheses are often contradictory, and need to be resolved by applying consensus methods. In this thesis, asymmetric median tree method (AMT) for construction of consensus phylogenetic trees is described, and along with two other approximation methods implemented in Biopython. AMT was compared to methods of strict, majority and Adams consensus trees on a number of artificial and one real data set. The comparisons were evaluated using the Robinson-Foulds and the tree resolution metrics. The results show that AMT is often the least similar to the input trees. At the same time, it is the most resolved tree, and therefore it offers the most information about evolutionary history of taxonomical units.

The increased rate of data collection, storage, and availability results in a corresponding interest for data analyses and predictive models based on simultaneous inclusion of multiple data sources. This tendency is ubiquitous in practical applications of machine learning, including recommender systems, social network analysis, finance and computational biology. The heterogeneity and size of the typical datasets calls for simultaneous dimensionality reduction and inference from multiple data sources in a single model. Matrix factorization and multiple kernel learning models are two general approaches that satisfy this goal. This work focuses on two specific goals, namely i) finding interpretable, non-overlapping (orthogonal) data representations through matrix factorization and ii) regression with multiple kernels through the low-rank approximation of the corresponding kernel matrices, providing non-linear outputs and interpretation of kernel selection. The motivation for the models and algorithms designed in this work stems from RNA biology and the rich complexity of protein-RNA interactions. Although the regulation of RNA fate happens at many levels - bringing in various possible data views - we show how different questions can be answered directly through constraints in the model design. We have developed an integrative orthogonality nonnegative matrix factorization (iONMF) to integrate multiple data sources and discover non-overlapping, class-specific RNA binding patterns of varying strengths. We show that the integration of multiple data sources improves the predictive accuracy of retrieval of RNA binding sites and report on a number of inferred protein-specific patterns, consistent with experimentally determined properties. A principled way to extend the linear models to non-linear settings are kernel methods. Multiple kernel learning enables modelling with different data views, but are limited by the quadratic computation and storage complexity of the kernel matrix. Considerable savings in time and memory can be expected if kernel approximation and multiple kernel learning are performed simultaneously. We present the Mklaren algorithm, which achieves this goal via Incomplete Cholesky Decomposition, where the selection of basis functions is based on Least-angle regression, resulting in linear complexity both in the number of data points and kernels. Considerable savings in approximation rank are observed when compared to general kernel matrix decompositions and comparable to methods specialized to particular kernel function families. The principal advantages of Mklaren are independence of kernel function form, robust inducing point selection and the ability to use different kernels in different regions of both continuous and discrete input spaces, such as numeric vector spaces, strings or trees, providing a platform for bioinformatics. In summary, we design novel models and algorithms based on matrix factorization and kernel learning, combining regression, insights into the domain of interest by identifying relevant patterns, kernels and inducing points, while scaling to millions of data points and data views. V času pospešenega zbiranja, organiziranja in dostopnosti podatkov se pojavlja potreba po razvoju napovednih modelov na osnovi hkratnega učenja iz več podatkovnih virov. Konkretni primeri uporabe obsegajo področja strojnega učenja, priporočilnih sistemov, socialnih omrežij, financ in računske biologije. Heterogenost in velikost tipičnih podatkovnih zbirk vodi razvoj postopkov za hkratno zmanjšanje velikosti (zgoščevanje) in sklepanje iz več virov podatkov v skupnem modelu. Matrična faktorizacija in jedrne metode (ang. kernel methods) sta dve splošni orodji, ki omogočata dosego navedenega cilja. Pričujoče delo se osredotoča na naslednja specifična cilja: i) iskanje interpretabilnih, neprekrivajočih predstavitev vzorcev v podatkih s pomočjo ortogonalne matrične faktorizacije in ii) nadzorovano hkratno faktorizacijo več jedrnih matrik, ki omogoča modeliranje nelinearnih odzivov in interpretacijo pomembnosti različnih podatkovnih virov. Motivacija za razvoj modelov in algoritmov v pričujočem delu izhaja iz RNA biologije in bogate kompleksnosti interakcij med proteini in RNA molekulami v celici. Čeprav se regulacija RNA dogaja na več različnih nivojih - kar vodi v več podatkovnih virov/pogledov - lahko veliko lastnosti regulacije odkrijemo s pomočjo omejitev v fazi modeliranja. V delu predstavimo postopek hkratne matrične faktorizacije z omejitvijo, da se posamezni vzorci v podatkih ne prekrivajo med seboj - so neodvisni oz. ortogonalni. V praksi to pomeni, da lahko odkrijemo različne, neprekrivajoče načine regulacije RNA s strani različnih proteinov. Z vzključitvijo več podatkovnih virov izboljšamo napovedno točnost pri napovedovanju potencialnih vezavnih mest posameznega RNA-vezavnega proteina. Vzorci, odkriti iz podatkov so primerljivi z eksperimentalno določenimi lastnostmi proteinov in obsegajo kratka zaporedja nukleotidov na RNA, kooperativno vezavo z drugimi proteini, RNA strukturnimi lastnostmi ter funkcijsko anotacijo. Klasične metode matrične faktorizacije tipično temeljijo na linearnih modelih podatkov. Jedrne metode so eden od načinov za razširitev modelov matrične faktorizacije za modeliranje nelinearnih odzivov. Učenje z več jedri (ang. Multiple kernel learning) omogoča učenje iz več podatkovnih virov, a je omejeno s kvadratno računsko zahtevnostjo v odvisnosti od števila primerov v podatkih. To omejitev odpravimo z ustreznimi približki pri izračunu jedrnih matrik (ang. kernel matrix). V ta namen izboljšamo obstoječe metode na način, da hkrati izračunamo aproksimacijo jedrnih matrik ter njihovo linearno kombinacijo, ki modelira podan tarčni odziv. To dosežemo z metodo Mklaren (ang. Multiple kernel learning based on Least-angle regression), ki je sestavljena iz Nepopolnega razcepa Choleskega in Regresije najmanjših kotov (ang. Least-angle regression). Načrt algoritma vodi v linearno časovno in prostorsko odvisnost tako glede na število primerov v podatkih kot tudi glede na število jedrnih funkcij. Osnovne prednosti postopka so poleg računske odvisnosti tudi splošnost oz. neodvisnost od uporabljenih jedrnih funkcij. Tako lahko uporabimo različne, splošne jedrne funkcije za modeliranje različnih delov prostora vhodnih podatkov, ki so lahko zvezni ali diskretni, npr. vektorski prostori, prostori nizov znakov in drugih podatkovnih struktur, kar je prikladno za uporabo v bioinformatiki. V delu tako razvijemo algoritme na osnovi hkratne matrične faktorizacije in jedrnih metod, obravnavnamo modele linearne in nelinearne regresije ter interpretacije podatkovne domene - odkrijemo pomembna jedra in primere podatkov, pri čemer je metode mogoče poganjati na milijonih podatkovnih primerov in virov.