M.Sc. / Please refer to full text to view abstract

The Support Vector Machine (SVM) classification method has recently gained popularity due to the ease of implementing non-linear separating surfaces. SVM is an optimization problem with the two competing goals, minimizing misclassification on training data and maximizing a margin defined by the normal vector of a learned separating surface. We develop and implement new SVM models based on previously conceived SVM with L_1-Norm regularization with ramp loss error terms. The goal being a new SVM model that is both robust to outliers due to ramp loss, while also easy to implement in open source and off the shelf mathematical programming solvers and relatively efficient in finding solutions due to the mixed linear-integer form of the model. To show the effectiveness of the models we compare results of ramp loss SVM with L_1-Norm and L_2-Norm regularization on human organ microbial data and simulated data sets with outliers.

Dans la première partie de cette thèse nous passons en revue la classification par modèle de mélange. En particulier nous décrivons une famille de mélanges gaussiens d’un usage courant, dont la parcimonie porte sur des paramètres d’interprétation géométrique. Comme ces modèles possèdent des inconvénients majeurs, nous leur opposons une nouvelle famille de mélanges dont la parcimonie porte sur des paramètres statistiques. Ces nouveaux modèles possèdent de nombreuses propriétés de stabilité qui les rendent mathématiquement cohérents et facilitent leur interprétation. Dans la seconde partie de ce travail nous présentons une méthode nouvelle dite de classification simultanée. Nous montrons que la classification d'un échantillon revient très souvent au partitionnement de plusieurs échantillons ; puis nous proposons d'établir un lien entre la population d'origine des différents échantillons. Ce lien, dont la nature varie selon le contexte, a toujours pour vocation de formaliser de façon réaliste une information commune aux données à classifier.Lorsque les échantillons sont décrits par des variables de même signification et que l'on cherche le même nombre de groupes dans chacun d'eux, nous établissons un lien stochastique entre populations conditionnelles. Lorsque les variables sont différentes mais sémantiquement proches d'un échantillon à l'autre, il se peut que leur pouvoir discriminant soit similaire et que l'imbrication des données conditionnelles soit comparable. Nous envisageons des mélanges spécifiques à ce contexte, liés par un chevauchement homogène de leurs composantes. / In the first part of this work we review the mixture model-based clustering method. In particular we describe a family of common Gaussian mixtures the parsimony of which is about geometrical parameters. As these models suffer from major drawbacks, we display new Gaussian mixtures the parsimony of which focuses on statistical parameters. These new models own many stability properties that make them mathematically consistent and facilitate their interpretation. In the second part of this work we display the so-called simultaneous clustering method. We highlight that the classification of a single sample can often be seen as a multiple sample clustering problem; then we propose to establish a link between the original population of the diverse samples. This link varies depending on the context but it always tries to formalize in a realistic way some common information of the samples to classify. When samples are described by variables with identical meaning and when the same number of groups is researched within each of them, we establish a stochastic link between the conditional populations. When the variables are different but semantically close through the diverse samples nevertheless their discriminant power may be similar and the nesting of the conditional data can be comparable. We consider specific mixtures dedicated to this context: the link between the populations consists in an homogeneous overlap of the components.

Thesis (Ed.M.)--Boston University / PLEASE NOTE: Boston University Libraries did not receive an Authorization To Manage form for this thesis or dissertation. It is therefore not openly accessible, though it may be available by request. If you are the author or principal advisor of this work and would like to request open access for it, please contact us at open-help@bu.edu. Thank you. / 2031-01-01

A research report submitted to the Faculty of Engineering and the Built Environment, University of the Witwatersrand, Johannesburg, in partial fulfilment of the requirements for the degree of Master of Science in Engineering. Johannesburg, 2017 / The study sought to establish the applicability of rock mass classification as a primary input to wall control blasting. Conventional rules of thumb are used to develop blast designs based on parametric ratios with insufficient consideration of the rock mass factors that influence the achievability of final wall designs. Control of the western highwall of the Phoenix pit had proven to be challenging in that the designed catchment berms and wall competence were perpetually unachievable from the pit crest to the current mining levels. This exposed the mining operation to safety hazards such as local wall rock failure from damaged crests, frozen toes and rolling rock falls from higher mining levels. There was also an effect of increased standoff distances from the concerned highwall which reduce the available manoeuvring area on the pit floor and subsequently the factor of extraction that is safely achievable. The study investigated the application of rock mass classification and the Blastability Index (BI) as a means to improve wall control. This was achieved by establishing zones according to rock type forming the western highwall rock mass wherein distinguishing rock mass classification factors were used to establish the suitable wall control designs through a Design Input Tool (DIT). The DIT consolidated rock mass classification methodologies such as the Geological Strength Index (GSI) and the Rock Mass Rating (RMR) and related them to the BI and discontinuities of the rock mass to produce a tool that can be used to develop objective wall control designs. The designs driven by the tool inherently take into account the rock mass characteristic factors at the centre of rock mass classification methods and significantly reduce the dependence on rule of thumb. It was found that this approach yields designs with powder factors that are consistent with the rock breaking effort and the behaviour of discontinuities while remaining biased towards preservation of perimeter wall rock. / MT 2017

Student Number : 0213068F - MSc Dissertation - School of Computer Science - Faculty of Science / Drug resistance of the HI virus due to its fast replication and error-prone mutation is a key factor in the failure to combat the HIV epidemic. For this reason, performing pre-therapy drug resistance testing and administering appropriate drugs or combination of drugs accordingly is very useful. There are two approaches to HIV drug resistance testing: phenotypic (clinical) and genotypic (based on the particular virus’s DNA). Genotyping tests HIV drug resistance by detecting specific mutations known to confer drug resistance. It is cheaper and can be computerised. However, it requires being able to know or learn what mutations confer drug resistance. Previous research using pattern recognition techniques has been promising, but the performance needs to be improved. It is also important for techniques that can quickly learn new rules when faced with new mutations or drugs. A relatively recent addition to these techniques is the Support Vector Machines (SVMs). SVMs have proved very successful in many benchmark applications such as face recognition, text recognition, and have also performed well in many computational biology problems where the number of features targeted is large compared to the number of available samples. This paper explores the use of SVMs in predicting the drug resistance of an HIV strain extracted from a patient based on the genetic sequence of those parts of the viral DNA encoding for the two enzymes, Reverse Transcriptase or Protease, which are critical for the replication of the HIV virus. In particular, it is the aim of this reseach to design the model without incorporating the biological knowledge at hand to enable the resulting classifier accommodate new drugs and mutations. To evaluate the performance of SVMs we used cross validation technique to measure the unbiased estimate on 2045 data points. The accuracy of classification and the area under the receiver operating characteristics curve (AUC) was used as a performance measure. Furthermore, to compare the performance of our SVMs model we also developed other prediction models based on popular classification algorithms, namely neural networks, decision trees and logistic regressions. The results show that SVMs are a highly successful classifier and out-perform other techniques with performance ranging between (94.13%–96.33%) accuracy and (81.26% - 97.49%) AUC. Decision trees were rated second and logistic regression performed the worst.

A research report submitted to the Faculty of Sciences, University of the Witwatersrand, Johannesburg in partial fulfilment of the requirements for the degree of Masters of Science in Environmental Sciences, 2017 / Modern tree biomass allometry makes use of “form factor”, which is the ratio of the true volume to the apparent volume. However, there is no database of form factors of South African trees, hence this study was undertaken to assess the possibility of assigning form factors to trees in a quick and easy way, either by visual assessment of an image of the tree or by simple field measurements. Stem diameter, taper and node length data for 112 trees was collected using both in situ and in-lab measurements from photos taken of the same trees in the field. The data were used to model tree volume using the fractal properties of branching architecture. The estimated tree volume was then used along with basal diameter and tree height to calculate the form factor. Results showed that measurements taken off images underestimated stem diameter and node length by 4% and 5% respectively, but the fractal allometry relationships developed using either the manual in-field or image analysis approach were not statistically different. This proves that dry season photography is sufficiently accurate for establishing relationships needed to construct a fractal model of tree volume. The image analysis approach requires a clear unobstructed view of the sample tree. This requirement made the approach less effective as when trees were in close proximity and when branches overlapped. The time taken using the photographic approach was twice the amount taken for the manual in-field. Form factor varied between species, but the variation was not statistically significant (p=0.579). The mean form factor per species ranged from 0.43 to 0.69. Form factors were negatively correlated with wood density (-0.177), basal diameter (-0.547) and height (-0.649). Due to the unavailability of an independent tree biomass dataset, it was impossible to validate the allometric equations based on estimated form factors and wood density. The inclusion of form factor was shown to improve the accuracy of biomass estimation by 11%. Principal component analysis showed that form factors can be assigned using tree height and the form quotient. / XL2018

High throughput sequencing data are rich in information and contain many off-target sequences (reads) that are often ignored but may be biologically relevant. Seed extension, a combination of reference and de novo based assembly methods, can be used to extract the information but it is time-consuming to implement because it requires that multiple seeds (sequences from one or many closely related species) be gathered in advance. A new tool is presented here, SeedSQrrL, that can automatically crawl the web to gather the seeds from the closest taxonomic relative for each gene and store it into a relational database. The seeds can then be used to create multiple seed extensions which are later combined into a reference or used for downstream phylogenetic analysis. Patterns in the resulting gene trees can be searched for using the traditional methods of tree comparison (Robinson-Foulds topological distance and branch-length comparison methods). Currently, no open source tree pattern matching program exists that allows the user to modify algorithms and create their own custom pattern matching functions. I have worked on such a tool, called Treematcher, and it will be made available in the ETE Toolkit (a Python Environment for Tree Exploration). Three biological case studies will be included included to demonstrate the capabilities of the two programs: 1) a custom function in Treematcher to perform a regular expression-like query, 2) SeedSQrrL will be used to isolate mitochondrial genes from snakes and chloroplast genes from angiosperms, and 3) a large case study of animals will be assembled. / A Dissertation submitted to the Department of Scientific Computing in partial fulfillment of the requirements for the degree of Doctor of Philosophy. / Spring Semester 2018. / April 2, 2018. / Automated Gene Reference Collection, Gene Tree Pattern Matching, High Throughput Sequence Analysis, NCBI Taxonomy, Open Source Software for Bioinformatics, Python / Includes bibliographical references. / Alan Lemmon, Professor Directing Dissertation; Michelle Arbeitman, University Representative; Anke Meyer-Baese, Committee Member; Peter Beerli, Committee Member; Dennis Slice, Committee Member.

Most classification rules can be expressed in terms of a distance (or dissimilarity) from the point to be classified to each of the candidate classes. For example, linear discriminant analysis classifies points into the class for which the (sample) Mahalanobis distance is smallest. However, dependence among these point-to-group distance measures is generally ignored. The primary goal of this project is to investigate the properties of a general non-parametric classification rule which takes this dependence structure into account. A review of classification procedures and applications is presented. The distance profile nearest-neighbor classification rule is defined. Properties of the rule are then explored via application to both real and simulated data and comparisons to other classification rules are discussed.

The Andes, along with the Amazon and Atlantic forests, harbor the richest avifauna in the world with roughly one third of all the world's species of birds. Many biogeographical studies have sought to explain the origin and diversification of Andean taxa. However, because of the Andes' extensive latitudinal span and complexity, there is no one single cause of origin or of diversification that can explain the diversity found in them. Along the Andes, multiple biogeographic patterns of disjunction between highland and lowland sister-groups have been linked to Andean uplift. For example, Ribas et al. (2007) provided evidence that the spatio-temporal diversification in the monophyletic parrot genus Pionus is causally linked to Andean tectonic and palaeoclimate change through vicariance. Thus, if the Andes uplift is responsible for some of the patterns of montane-lowland disjunctions, it may be one of the mechanisms underlying the taxonomic assembly of the Andean montane avifauna. In this dissertation I explored whether the origin and diversification of three groups of Andean birds--the exclusively Andean parrot genera Hapalopsittaca, the subclade of mangoes containing Doryfera, Schistes, and Colibri, and the ovenbirds of the tribe Thripophagini--can be linked to Earth history. The results show that the origin of these Andean taxa can be explained through vicariance from their lowland sister-groups, mediated by the uplift of the Andes. Thus, this thesis proposes that geological events are directly responsible for originating diversity throughout montane environments. Once in the Andes, the diversification of these montane taxa can be explained by events such as the tectonic evolution of the Andes--which created canyons and valleys that may have caused the vicariance of continuous populations--as well as by the climatic oscillation of the Pleistocene, which caused altitudinal shifts, expansion, and contraction of the montane vegetation belts during the climatic oscillations of the Pleistocene.