LEE- Matematik Mühendisliği-Doktora
Bu koleksiyon için kalıcı URI
Gözat
Son Başvurular
1 - 5 / 8
-
ÖgeTopological data analysis and clustering algorithms in machine learning(Graduate School, 2023-03-13)In this dissertation, we define a new non-Archimedean metric (a.k.a. an ultra-metric) called cophenetic metric on persistent homology classes of all degrees using only homological information. Then, based on numerical experiments on different datasets, we statistically verify that the topological information coming from the zeroth persistent homology with our cophenetic metric is consistent with the information provided by different hierarchical clustering algorithms using different metrics. We also observe that the clusters we obtained via the cophenetic metric do yield competitive silhouette scores and the Rand indices in comparison with clusters obtained from other metrics. The homological information about a filtered simplicial complex over the poset of positive real numbers is often presented by a barcode which depicts the evolution of the associated Betti numbers. However, there is wonderfully complex combinatorics associated with the homology classes of a filtered complex, and one can do more than just count them over the index poset. In this thesis, we show that this combinatorial information can be encoded by a filtered matroid, or even better, by rooted forests. We also show that these rooted forests can be realized as cobordisms. The subject of this thesis is presented in Chapter 1, in which we also provide a survey of the literature and a brief introduction to topological data analysis (TDA). In Chapter 2, we summarize the fundamental concepts from topological and metric spaces that we will need in subsequent chapters. In Chapter 3, we will go more deeply into the mathematical foundations of clustering techniques. Clustering algorithms, especially hierarchical clustering algorithms, play a key role in this dissertation. Comparisons of various clustering strategies also play important roles in this thesis. For this reason, we also survey numerical metrics to assess such clustering strategies in the same chapter. We also work out a complete example on a collection of Turkish cities and the distances between them, and compare the resulting clusters. The TDA calculations we make in this thesis heavily use simplicial and chain complexes, and their homologies. In all examples, we first calculate the homology of a filtered simplicial complexes in order to calculate its persistent homology. The primary computation procedure begins by distilling a point cloud into a filtered simplicial complex, followed by calculating the homology of the resulting filtered differential graded complex. Chapters 4 and 5 provide all the necessary background information on simplicial complexes, differential graded complexes, and homologies. We also present complete worked-out examples of calculating homology classes for a collection of simple simplicial complexes. To capture the exact topological features of the point cloud, the main challenge is to find an optimal proximity parameter for the simplicial technology. Persistent homology is developed precisely to deal with this issue. It keeps track of how long each topological feature of the supplied data endures as the proximity parameter varies. It produces multisets of intervals represented as barcodes. We cover these concepts in Chapter 6. Chapter 7 is devoted to developing a rigorous theoretical foundation for filtered simplicial and chain complexes. Along with developing such a foundation for filtered complexes, we also discovered an exciting combinatorial representation for homologies of filtered complexes in the form of filtered matroids, especially by employing dendrograms labeled by circuits in a matroid. In Chapter 8, we defined a new non-archimedean metric called homological cophenetic distance on homology classes which is the main contribution of this thesis to literature. In Chapter 9 we are concerned with developing another presentation of the homological information coming from a filtered complex using cobordisms of punctured spheres, which are themselves punctured higher dimensional spheres. In Chapter 10, we give a detailed account of how we used the cophenetic distance in our numerical experiments. In the first experiment, we compared the standard zeroth persistent homology representations (barcodes), the dendrogram we derived from the cophenetic distance, and the dendrogram coming from the hierarchical clustering algorithm with the Euclidean metric on a small synthetic dataset embedded in $\mathbb{R}^2$. In the second experiment, we studied the geographic coordinates of a small sample of Turkish cities. By following the same statistical comparison methodology as in the first experiment, we compare the dendrograms produced by cophenetic metrics on the zeroth homology and the dendrograms produced by hierarchical clustering algorithms using a variety of distance measurements in order to assess the validity of our study. In the third and last experiment, we applied the hierarchical clustering algorithms to various datasets (the dataset of Turkish cities, the Iris dataset, the Cancer Coimbra dataset, and two synthetic datasets) with different metrics including our cophenetic metric, and we statistically analyzed the clustering results. Finally, in Chapter11, we summarize and analyze the results we obtain, and the constraints of our approach such as need for high computational power and high memory requirements. We also discuss potential future research directions one can follow to extend this thesis such as designing suitable applications with real-world datasets, visualizing cobordisms, or developing filtered combinatorial simplicial complexes on a categorical dataset.
-
ÖgeTime series classification via topological data analysis(Graduate School, 2022-06-22)This dissertation aims to demonstrate the power of Topological Data Analysis (TDA) and the subwindowing method for feature engineering in time series classification tasks. As an application, we used two publicly available datasets, WESAD and DriveDB. These datasets consisted of physiological signals collected under stressful and non stressful events. Furthermore, in order to assess the reliability of our methodology, we tested our feature engineering methods on a synthetic dataset that consists of artificial physiological signals mimicking a stress detection study. The results indicated that automatically created topological features can yield higher classification accuracies than signal-specific and hand-crafted features (such as heart rate derived from an ECG signal). In the first chapter of this work, we briefly summarize TDA and persistent homology. Also, the methods for time series classification via persistent homology is discussed, and we make a literature review on the subject. The second chapter is devoted to time series methods, and how we can classify them. We first define the method of sliding windows, and discuss why it can be useful in machine learning tasks. Then, we talk about time delay embeddings which transforms a univariate time series into a high dimensional dataset. We illustrate how the topology of the resulting dataset is affected by the delay parameter (also known as the embedding dimension). At the end of this chapter, we introduced the subwindowing methodology which solved the main problem of this work. We showed that this method allows us to reduce noise, improve computation time by a large amount, and use longer windows without incurring extra computational cost. In the third chapter, the theoretical background for TDA and persistent homology is given. The chapter starts with discussing why we should see the data at different scales. Then we give preliminary definitions related with simplices and simplicial complexes. We state the Nerve theorem and talk about how a topological space and a simplicial complex can be homotopy equivalent under some assumptions. This theorem tells us that the Cech (and therefore Rips) complexes are topologically similar to the underlying object that the dataset was sampled from. Later in this chapter, we define simplicial homology and show how we can compute the homology of a simplicial complex. Note that we need a fixed distance (epsilon) parameter to build a simplicial complex on top of a dataset. On the other hand, persistent homology allows us to investigate the persistence of homology groups when epsilon varies. After presenting how persistent homology works, we define persistence diagrams and two widely used metrics between them. Also, we show that persistence diagrams are stable under small perturbations of the data. Lastly, we show some means of performing feature engineering of persistence diagrams. The fourth chapter consists of the description of the datasets used in this dissertation and our methodologies. First, we introduce the three datasets (synthetic, WESAD and DriveDB) used in this study. For the synthetic dataset, there were two classes of physiological signals: stress and non stress. The classes for WESAD were baseline, amusement and stress. For DriveDB, the classes were relax, driving in the highway (low stress) and driving in the city (high stress). We then talk about the physiological signals included in the datasets, their sampling frequencies, and some preprocessing we did beforehand. Our experiments had some parameters such as window size, subwindow size, the embedding dimension in time delay embeddings. Later in this chapter, we discuss how these parameters were chosen, and how we did feature engineering for our experiments. Then, we present the machine learning algorithms and their hyperparameters used in our experiments. Lastly in chapter four, we introduce the two cross-validation methodologies used in our experiments. For Leave-one-subject-out cross validation (LOSOCV), the model is trained on all subjects but one, and tested on the other. When each participant appears in the test set once and only once, the results are averaged. This cross validation technique tells us about the model's performance on a previously unseen subject. For intra-subject cross validation, we split each subjects data into two. We train on either half, and test on the other, then average the results. We get a final accuracy by averaging all accuracies obtained from each participant. This method shows whether the model can benefit from having the same subject's data on both the train and the test sets. The results of the experiments are covered in the fifth chapter. We presented the results for the synthetic, WESAD and DriveDB datasets, respectively. The results for the synthetic dataset indicated that as the magnitude of the physiological change that mimics stress increases, stress detection accuracy also improves. For example, when the heart rate variability -an important stress indicator- is raised, the topological features could detect it almost perfectly. The results imply that stress detection errors in real-world datasets can be attributed to the noisy nature of the dataset itself, rather than the topological features. For example, such lack of effect can appear when some participants do not react to the stress condition. When the results from the real datasets were investigated, we usually observed the highest affect recognition accuracies when features coming from all persistence diagrams (level sets and delay embeddings) are used. Nevertheless, using only one persistence diagram (resulting in much fewer features) we were able to achieve similar recognition performance. This tells us that high accuracies are attainable using a small number of automatically engineered topological features rather than hand-crafted signal-specific features. For the three-class tasks, we observed that stress conditions are well separated from other conditions. This result supports the hypothesis that topological features works pretty well in distinguishing chaotic time series from non chaotic ones. When we made a binary classification task (stress vs non stress), topological features again performed better than those used in the original studies for most of the physiological signals. We have already stated that an important advantage of the subwindowing method is to be able to change the window size effectively. When we tested different window sizes, we observed that higher windows implied better stress detection performance. Furthermore, model performance with intra-subjects cross validation was significantly higher than LOSOCV. This was an expected finding since the model can perform better on the test set when the data from the same subject appears in the training set. Finally, in the sixth chapter, we outline our methodologies and their limitations. We also discussed what future works can aim for. For example, future studies can assess the performance of a model trained on one dataset and tested on another. Also, later research can use semi-supervised (rather than supervised) tasks for even improved accuracies. Lastly, one can use other vector representations of persistence diagrams for feature engineering.
-
ÖgeParameter optimization for mathematical modeling(Graduate School, 2023-06-09)Mathematical modeling is used to explain and forecast complex systems, and parameter optimization methods have a crucial role to find the optimal set of parameters obtained by minimizing an objective function. Also, the management of computational resources is essential for handling big models in real-time scenarios. A. Duran and G. Caginalp (2008) propose a hybrid parameter optimization forecast algorithm for asset prices via asset flow differential equations. In this thesis, we propose a new mathematical method for an inverse problem of parameter vector optimization in asset flow theory. For this purpose, we use quasi-Newton (QN) and Monte Carlo simulations to optimize the function F[K] for each selected event and initial parameter vector. We present grid and random methods and conclude that the grid approach is better than the random approach in the unconstrained optimization problem. This study also presents a parallel numerical parameter optimization algorithm for dynamical systems used in financial applications. It achieves speed-up for up to 512 cores and considers more extensive financial market situations. Moreover, it also evaluates the convergence of the model parameter vector via nonlinear least squares error, and maximum improvement factor. In this thesis, we also examine the performance, scalability, and robustness of OpenFOAM on the GPGPU cluster for bio-medical fluid flow simulations. It compared the CPU performance of iterative solver icoFoam with direct solver SuperLU_DIST 4.0 and hybrid parallel codes of MPI+OpenMP+CUDA versus MPI+OpenMP implementation of SuperLU_DIST 4.0. Results showed speed-up for large matrices up to 20 million x 20 million. Besides that, we investigate the usage of eigenvalues to examine the spectral effects of large matrices on the performance of scalable direct solvers. Gerschgorin's theorem can be used to bound the spectrum of square matrices, and behaviors such as disjoint, overlapped, or clustered Gerschgorin circles can give clues. We define the minimum number of cores and show that it depends on the sparsity level and size of the matrix, increasing slightly as the sparsity level decreases and the order increases. In sum, this thesis presents new methods for initial parameter selection and a new algorithm for parallel numerical parameter optimization. Also, we define new metrics and show that the importance of right matching for computational systems and the optimal minimum number of cores are important in mathematical modeling and simulation.
-
ÖgeManifolds of generalised G-structures in string compactifications(Graduate School, 2023-03-22)A G-structure on a differentiable manifold M of dimension n can be described as a reduction of the linear frame bundle L(M) of M to a Lie subgroup G of $GL(n,\mathbb{R})$. Such a reduction is equivalent to the existence of certain geometric structures on M, depending on what the subgroup G is. For example, an O(n)-structure corresponds to the existence of a Riemannian metric g. Similarly, by the existence of an almost complex structure J, the structure group reduces to $GL(n/2,\mathbb{C})$. If a Riemannian metric and an almost complex structure are compatible and the metric is hermitian then the structure group reduces to SU(n/2). In a similar fashion, a generalized G-structure can be described as a reduction of the structure group of the principal bundle associated with the generalized tangent bundle $TM\oplus T^*M$. The natural structure group of $TM\oplus T^*M$ is O(n,n). The generalized G-structures also correspond to the existence of certain geometrical objects. For example, the reduction of the structure group from O(n,n) to $O(n)\times O(n)$ corresponds to the existence of a generalized metric. Similarly, on an even-dimensional real manifold $M$ a generalized almost complex structure is given by a reduction of the structure group from O(n,n) to U(n/2,n/2). A generalized almost complex structure is defined by the existence of a pure spinor which is a section of the exterior bundle $\bigwedge^\bullet T^* M$. The SU(n/2,n/2)-structure is equivalent to the existence of a globally defined pure spinor of non-vanishing norm. Furthermore, $SU(n/2)\times SU(n/2)$-structure is given by the existence of two compatible pure spinors. The main theme of this thesis is the study of manifolds of generalized G-structure relevant to string compactifications. Superstring theory is a quantum theory of gravity consistent in 10 dimensions. There are five consistent superstring theories and the low energy dynamics of massless space-time fields are governed by ten-dimensional supergravity theories. The supergravity field equations are nonlinear partial differential equations that can be regarded as a generalization of field equations of Einstein's theory of general relativity (GR). In a supersymmetric compactification of Type II string theory down to 4 dimensions, it is required that the structure group of the generalized tangent bundle $TM \oplus T^*M$ of the six-dimensional internal manifold M is reduced from SO(6,6) to $SU(3) \times SU(3)$. This is equivalent to the existence of two globally defined compatible pure spinors $\Phi_1$ and $\Phi_2$. Furthermore, these pure spinors should satisfy certain first-order differential equations, namely supersymmetry equations. We show that these equations are covariant under certain Pin(d,d) transformations. We also show that Non-Abelian T-duality (NATD) which is generated by a coordinate-dependent Pin(d,d) transformation is a particular solution generating transformation for these pure spinor equations. Our method is demonstrated by studying the NATD of a specific class of geometries with SU(2) isometry and SU(3)-structure. Some of the manifolds belonging to this class are $AdS_5\times T^{1,1}$, $AdS_5\times Y^{p,q}$ and $AdS_5\times S^5$. It is interesting to note that in each case, the internal manifold is a Sasaki-Einstein manifold. We show that the transformed pure spinors are associated with an SU(2)-structure. The plan of the thesis is as follows: in section 2, we study principal fiber bundles, vector bundles, and linear frame bundles. Then, we study the concept of the reduction of the structure groups. We also give familiar examples of G-structures in detail. In section 3, we briefly review the relation between G-holonomy and torsion-free G-structures. In section 4, we study the basic concepts regarding the geometry of the generalized tangent bundle $TM\oplus T^*M$. This leads us to the definition of a generalized G-structure. Since our main interest is in $SU(3)\times SU(3)$-structures we give in a separate subsection the description of $SU(3)\times SU(3)$-structures and the associated pure spinors in detail. In section 5, we focus on the differential equations to be satisfied by the pure spinors for preservation of ${\cal{N}}=1$ supersymmetry. We study the covariance of these equations under constant and non-constant Pin(d,d) transformations. Then, we study Non-Abelian T-duality (NATD) transformations in detail, and we show the invariance of pure spinor equations under NATD. In section 6, we consider a specific class of geometries. We transform the pure spinors associated with the SU(3)-structure and show that the resulting pure spinors determine an SU(2) structure. We also study the NATD transformation of the metric, the B field, and the Ramond-Ramond fields.
-
ÖgeThe generalized fractional Benjamin Bona Mahony equation: Analytical and numerical results(Lisansüstü Eğitim Enstitüsü, 2021)In this thesis study we consider the generalized fractional Benjamin-Bona-Mahony (gfBBM) equation u_t+ u_x + \frac{1}{2}(u^{p+1})_x+ \frac{3}{4}D^{\alpha} u_{x}+ \frac{5}{4}D^{\alpha} u_{t}=0, where $x$ and $t$ represents spatial coordinate and time, respectively. This equation is derived to model the propagation of small amplitude long unidirectional waves in a nonlocally and nonlinearly elastic medium. The gfBBM equation has a general power-type nonlinearity and two fractional-type terms. Thanks to these properties, the gfBBM equation is noticed as a satisfactory and interesting model in the literature. The aim of this thesis study is to perform various mathematical and numerical analyses for the gfBBM equation and to understand the influence of nonlinearity and fractional dispersion on the dynamics of solutions. The thesis study is organized in the following way: In the first chapter, we briefly introduce the general background on the fractional type nonlinear partial differential equations with lower dispersion such as fractional Korteweg de Vries (fKdV) and fractional Benjamin-Bona-Mahony (fBBM) and gfBBM equations. Then, we propose derivation and some properties of the gfBBM equation. We also state the analytical and numerical methods used to solve this equation. Furthermore, the literature overview on gfBBM and related equations is given in this chapter. The second chapter is devoted to the analytical results for the gfBBM equation. In the first section of this chapter we recall the preliminaries. This section contains useful definitions related to functional analysis, lemmas and theorems used in the thesis. In the second section, we derive conserved quantities of the gfBBM equation. We also find constraints on the order $\alpha$ of the fractional term. The aim of the third section is to prove the local well-posedness of the Cauchy problem for the gfBBM equation together with the initial condition u(x,0)=u_0 (x). For the case $1 \leq \alpha \leq 2$, we prove the local well-posedness of the solutions by using contraction mapping principle. On the other hand, for the case $0 < \alpha < 1$, we use the approaches given for the fBBM equation by He and Mammeri (2018). Therefore, we consider the regularization of the Cauchy problem for the gfBBM equation and then use the convergence of regularized solutions to the solutions of main problem. The section 4 presents the conditions for the non-existence of solitary wave solutions to the gfBBM equation. Existence and uniqueness of solitary wave solutions are obtained by using the result of Frank and Lenzmann (2013). We also consider the restrictions on the $\alpha$ and speed of wave $c$ so that the gfBBM equation admits positive or negative solitary waves. Finally, we derive exact solitary wave solutions to the gfBBM equation for the special cases $\alpha=1$ and $\alpha=2$ when $p=1$. In the last section of this chapter we discuss the stability properties of solitary wave solutions associated to the gfBBM equation. We first give the Hamiltonian formulation of the equation. Then, we prove the orbital stability of solitary wave solutions by using approach given by Grillakis Shatah Strauss (GSS) (1987) and for the stability we obtain following conditions when $1 \leq p \leq 4$: 1) $\frac{p}{p+2}<\alpha < \frac{p}{2}$ and $c>c_{1,p}>1$, 2) $\frac{p}{2}<\alpha < 2$ and $c>1$ or $\frac{3}{5}>c>c_{2,p}$, with $c_{1,p}=\frac{6\alpha + 2p + 3 \alpha p + \sqrt 2 p \sqrt{2 \alpha - p + \alpha p} }{5(2 \alpha + \alpha p)}$ and $c_{2,p}=\frac{6\alpha + 2p + 3 \alpha p - \sqrt 2 p \sqrt{2 \alpha - p + \alpha p} }{5(2 \alpha + \alpha p)}$. In the last chapter, we present the numerical results for the gfBBM equation. We first state efficient numerical algorithms for gfBBM equation and then carry out various numerical experiments. The Petviashvili method is proposed for the generation of the solitary wave solutions that cannot be obtained analytically. We numerically investigate the effects of the relation between the nonlinearity and the dispersion on the solutions. The evolution of generated wave profiles in time is investigated numerically by Fourier pseudo-spectral method. The efficiency of the methods will be demonstrated by various numerical simulations.