Doctor of Philosophy / Computing and Information Sciences / Doina Caragea / Increasing amounts of content on the Web means that users can select from a wide variety of items (i.e., items that concur with their tastes and requirements). The generation of personalized item suggestions to users has become a crucial functionality for many web applications as users benefit from being shown only items of potential interest to them. One popular solution to creating personalized item suggestions to users is recommender systems. Recommender systems can address the item recommendation task by utilizing past user preferences for items captured as either explicit or implicit user feedback. Numerous collaborative filtering (CF) approaches have been proposed in the literature to address the recommendation problem in the single-domain setting (user preferences from only one domain are used to recommend items). However, increasingly large datasets often prevent experimentation of every approach in order to choose the one that best fits an application domain. The work in this dissertation on the single-domain setting studies two CF algorithms, Adsorption and Matrix Factorization (MF), considered to be state-of-the-art approaches for implicit feedback and suggests that characteristics of a domain (e.g., close connections versus loose connections among users) or characteristics of data available (e.g., density of the feedback matrix) can be useful in selecting the most suitable CF approach to use for a particular recommendation problem. Furthermore, for Adsorption, a neighborhood-based approach, this work studies several ways to construct user neighborhoods based on similarity functions and on community detection approaches, and suggests that domain and data characteristics can also be useful in selecting the neighborhood approach to use for Adsorption. Finally, motivated by the need to decrease computational costs of recommendation algorithms, this work studies the effectiveness of using short-user histories and suggests that short-user histories can successfully replace long-user histories for recommendation tasks. Although most approaches for recommender systems use user preferences from only one domain, in many applications, user interests span items of various types (e.g., artists and tags). Each recommendation problem (e.g., recommending artists to users or recommending tags to users) can be considered unique domains, and user preferences from several domains can be used to improve accuracy in one domain, an area of research known as cross-domain recommender systems. The work in this dissertation on cross-domain recommender systems investigates several limitations of existing approaches and proposes three novel approaches (two Adsorption-based and one MF-based) to improve recommendation accuracy in one domain by leveraging knowledge from multiple domains with implicit feedback. The first approach performs aggregation of neighborhoods (WAN) from the source and target domains, and the neighborhoods are used with Adsorption to recommend target items. The second approach performs aggregation of target recommendations (WAR) from Adsorption computed using neighborhoods from the source and target domains. The third approach integrates latent user factors from source domains into the target through a regularized latent factor model (CIMF). Experimental results on six target recommendation tasks from two real-world applications suggest that the proposed approaches effectively improve target recommendation accuracy as compared to single-domain CF approaches and successfully utilize varying amounts of user overlap between source and target domains. Furthermore, under the assumption that tuning may not be possible for large recommendation problems, this work proposes an approach to calculate knowledge aggregation weights based on network alignment for WAN and WAR approaches, and results show the usefulness of the proposed solution. The results also suggest that the WAN and WAR approaches effectively address the cold-start user problem in the target domain.

abstract: Modern measurement schemes for linear dynamical systems are typically designed so that different sensors can be scheduled to be used at each time step. To determine which sensors to use, various metrics have been suggested. One possible such metric is the observability of the system. Observability is a binary condition determining whether a finite number of measurements suffice to recover the initial state. However to employ observability for sensor scheduling, the binary definition needs to be expanded so that one can measure how observable a system is with a particular measurement scheme, i.e. one needs a metric of observability. Most methods utilizing an observability metric are about sensor selection and not for sensor scheduling. In this dissertation we present a new approach to utilize the observability for sensor scheduling by employing the condition number of the observability matrix as the metric and using column subset selection to create an algorithm to choose which sensors to use at each time step. To this end we use a rank revealing QR factorization algorithm to select sensors. Several numerical experiments are used to demonstrate the performance of the proposed scheme. / Dissertation/Thesis / Doctoral Dissertation Applied Mathematics 2015

Coordenação de Aperfeiçoamento de Pessoal de Nível Superior / When the controller s performance is reduced along the time and the control efficiency is no longer satisfactory, then it is necessary to develop certain techniques for continuously monitor control systems such as it can keep track of process changes maintaining the characteristics of efficiency of the original controller design. The divergences between controlled variables and the desired setpoint can be quantified by a single number that is known as performance index. In the present work, a novel way of evaluating the control process performance is introduced and applied as a new measure of performance for a control structure. For doing that, the knowledge of the invariants of the process under feedback and factorization tools of linear local models were applied for designing a new reference model. The proposed techniques were used as a reference of what can or cannot be achieved by a particular system operating under the action of a feedback control law: minimum variance control (MVC), proportional-integral (PI), and model predictive control (MPC). Hence, the proposed performance indexes are the tools for the evaluation of the behavior of the control system during regular industrial operation, not only based upon the physical point of view, in which the reference is treated as a realizable system, but also from a theoretical point of view, in which the model factorization procedure takes place, and therefore allowing the extraction of the invariant part of a system, changing the system to its easiest controlled form description, and then serving as a reference model for the calculation of the performance index, with a low computing cost and preserving the importance of interactions and high frequency effects for balancing the reference behavior. In this work was investigated, with success, the effect of coupling, interactions and quality degradation of the control for SISO (Single-Input and Signle-Output) and MIMO (Multiple-Input and Multiple-Output) processes. The results show that the proposed technique are effective and a viable tool for process monitoring and controlled systems performance evaluation. / Quando o desempenho de um controlador se reduz ao longo do tempo e a eficiência de controle já não é mais satisfatória, tem-se que o controle em vigor ficou obsoleto devido às mudanças ocorridas no processo. Dessa forma, surge a necessidade de se desenvolver técnicas de monitoramento contínuo de sistemas de controle para que o mesmo possa acompanhar as mudanças ocorridas no processo, mantendo as características de eficiência do projeto original do controlador. As divergências entre variáveis controladas e os valores desejados (setpoint) podem ser quantificadas por um número que ´e conhecido como ´índice de desempenho. Neste trabalho, investiga-se o desempenho de sistemas de controle através de novas medidas de desempenho para estruturas de controle. As medidas propostas utilizam o conhecimento dos invariantes do processo obtidos através de ferramentas de fatoração de modelos locais lineares e da definição de um padrão de referencia `a ação possível de controle para plantas operando com controladores feedback para casos de controladores com variância mínima (MVC), proporcional-integral (PI) e controlador preditivo baseado em modelos (MPC). Desta forma, a técnica proposta neste trabalho serve como uma ferramenta para a avaliação do comportamento do sistema de controle durante a operação industrial, pois, tanto do ponto de vista físico, por se tratar de uma referencia realizável, quanto do ponto de vista teórico, que ao realizar a fatoração, as informações dos invariantes dos modelos são extraídos, fazendo com que os mesmos se tornem mais fáceis de serem controlados, então, servindo como referencia para o cálculo do ´índice de desempenho para o sistema. As técnicas propostas possuem baixo custo computacional e preservam a importância de interações e efeitos de elevadas freqüências no compromisso entre a atenuação da sensibilidade das perturbações e a velocidade de resposta da malha de controle avaliada. Investigou-se então, com sucesso, o efeito de acoplamentos, interações e degradação da qualidade do controle para processos com uma entrada e uma saída (SISO) e sistemas com múltiplas entradas e múltiplas saídas (MIMO), comprovando então a eficácia da técnica proposta como uma ferramenta viável para o monitoramento e avaliação de sistemas de controle. / Mestre em Engenharia Química

abstract: The problem of systematically designing a control system continues to remain a subject of intense research. In this thesis, a very powerful control system design environment for Linear Time-Invariant (LTI) Multiple-Input Multiple-Output (MIMO) plants is presented. The environment has been designed to address a broad set of closed loop metrics and constraints; e.g. weighted H-infinity closed loop performance subject to closed loop frequency and/or time domain constraints (e.g. peak frequency response, peak overshoot, peak controls, etc.). The general problem considered - a generalized weighted mixed-sensitivity problem subject to constraints - permits designers to directly address and tradeoff multivariable properties at distinct loop breaking points; e.g. at plant outputs and at plant inputs. As such, the environment is particularly powerful for (poorly conditioned) multivariable plants. The Youla parameterization is used to parameterize the set of all stabilizing LTI proper controllers. This is used to convexify the general problem being addressed. Several bases are used to turn the resulting infinite-dimensional problem into a finite-dimensional problem for which there exist many efficient convex optimization algorithms. A simple cutting plane algorithm is used within the environment. Academic and physical examples are presented to illustrate the utility of the environment. / Dissertation/Thesis / M.S. Electrical Engineering 2013

Cette thèse est consacrée au démixage en imagerie hyperspectrale en particulier dans le cas où des objets de petite taille sont présents dans la scène. Les images hyperspectrales contiennent une grande quantité d’information à la fois spectrale et spatiale, et chaque pixel peut être vu comme le spectre de réflexion de la zone imagée. Du fait de la faible résolution spatiale des capteurs le spectre de réflexion observé au niveau de chaque pixel est un mélange des spectres de réflexion de l’ensemble des composants imagés dans le pixel. Une problématique de ces images hyperspectrales est le démixage, qui consiste à décomposer l’image en une liste de spectres sources, appelés endmembers, correspondants aux spectres de réflexions des composants de la scène d’une part, et d’autre part la proportion de chacun de ces spectres source dans chaque pixel de l’image. De nombreuses méthodes de démixage existent mais leur efficacité reste amoindrie en présence de spectres sources dits rares (c’est-à-dire des spectres présents dans très peu de pixels, et souvent à un niveau subpixelique). Ces spectres rares correspondent à des composants présents en faibles quantités dans la scène et peuvent être vus comme des anomalies dont la détection est souvent cruciale pour certaines applications.Nous présentons dans un premier temps deux méthodes de détection des pixels rares dans une image, la première basée sur un seuillage de l’erreur de reconstruction après estimation des endmembers abondants, la seconde basée sur les coefficients de détails obtenus par la décomposition en ondelettes. Nous proposons ensuite une méthode de démixage adaptée au cas où une partie des endmembers sont connus a priori et montrons que cette méthode utilisée avec les méthodes de détection proposées permet le démixage des endmembers des pixels rares. Enfin nous étudions une méthode de rééchantillonnage basée sur la méthode du bootstrap pour amplifier le rôle de ces pixels rares et proposer des méthodes de démixage en présence d’objets de petite taille. / This thesis is devoted to the unmixing issue in hyperspectral images, especiallyin presence of small sized objects. Hyperspectral images contains an importantamount of both spectral and spatial information. Each pixel of the image canbe assimilated to the reflection spectra of the imaged scene. Due to sensors’ lowspatial resolution, the observed spectra are a mixture of the reflection spectraof the different materials present in the pixel. The unmixing issue consists inestimating those materials’ spectra, called endmembers, and their correspondingabundances in each pixel. Numerous unmixing methods have been proposed butthey fail when an endmembers is rare (that is to say an endmember present inonly a few of the pixels). We call rare pixels, pixels containing those endmembers.The presence of those rare endmembers can be seen as anomalies that we want todetect and unmix. In a first time, we present two detection methods to retrievethis anomalies. The first one use a thresholding criterion on the reconstructionerror from estimated dominant endmembers. The second one, is based on wavelettransform. Then we propose an unmixing method adapted when some endmembersare known a priori. This method is then used with the presented detectionmethod to propose an algorithm to unmix the rare pixels’ endmembers. Finally,we study the application of bootstrap resampling method to artificially upsamplerare pixels and propose unmixing methods in presence of small sized targets.

Com o avanço da internet, novos tipos de negócios surgiram. Por exemplo, o sistema de anúncios online: produtores de sites e diversos outros conteúdos podem dedicar em uma parte qualquer de sua página um espaço para a impressão de anúncios de diversas lojas em troca de um valor oferecido pelo anunciante. É neste contexto que este trabalho se insere. O objetivo principal é o desenvolvimento de algoritmos que preveem a probabilidade que um dado usuário tem de se interessar e clicar em um anúncio a que está sendo exposto. Este problema é conhecido como predição de CTR (do inglês, \"Click-Through Rate\") ou taxa de conversão. Utiliza-se para isto uma abordagem baseada em regressão logística integrada a técnicas de fatoração de matriz que preveem, através da obtenção de fatores latentes do problema, a probabilidade de conversão para um anúncio impresso em dado site. Além disto, testes considerando uma estratégia dinâmica (em função do tempo) são apresentados indicando que o desempenho previamente obtido pode melhorar ainda mais. De acordo com o conhecimento do autor, esta é a primeira vez que este procedimento é relatado na literatura. / With the popularization of the internet, new types of business are emerging. An example is the online marketing system: publishers can dedicate in any given space of theirs websites a place to the printing of banners from different stores in exchange for a fee paid by the advertiser. It\'s in this context that this work takes place. Its main goal will be the development of algorithms that forecasts the probability that a given user will get interested in the ad he or she is seeing and click it. This problem is also known as CTR Prediction Task. To do so, a logistic regression approach is used combined with matrix factorization techniques that predict, through latent factor models, the probability that the click will occur. On top of that, several tests are conducted utilizing a dynamic approach (varying in function of time) revealing that the performance can increase even higher. According to the authors knowledge, this is the first time this test is conducted on the literature of CTR prediction.

Orientador: Maria Zoraide Martins Costa Soares / Dissertação (mestrado profissional) - Universidade Estadual de Campinas, Instituto de Matematica, Estatistica e Computação Cientifica / Made available in DSpace on 2018-08-10T23:10:07Z (GMT). No. of bitstreams: 1 Campos_LudioEdsondaSilva_M.pdf: 1007258 bytes, checksum: 78308663e7f18b51bedeb284cadca66a (MD5) Previous issue date: 2008 / Resumo: Neste trabalho abordamos algumas fatorações de matrizes, com vistas à resolução de sistemas lineares através de métodos diretos. Enfocamos particularmente as decomposições LU, Cholesky e QR, cujo uso tem sido largamente difundido em implementações computacionais. Nosso objetivo é apresentar um texto didático, acessível a alunos de graduação, que contemple a teoria básica de cada fatoração, incluindo a demonstração dos principais resultados, e que também forneça condições para uma primeira implementação de cada decomposição. Sugerimos alguns algoritmos, que foram implementados no software livre OCTAVE, através dos quais comparamos o tempo gasto para resolução de alguns sistemas lineares, utilizando as fatorações citadas / Abstract: In this work we discuss some matrix factorizations, with a view to the resolution of linear systems through direct methods. We focus particularly the LU, Cholesky and QR decompositions, whose use has been widely spread in computer implementations. Our goal is to present a didactic text, accessible to undergraduate students, which contemplates the basic theory of each factorization, including the demonstration of the main result and that also provide conditions for a first implementation of each decomposition. We suggest some algorithms that were scheduled in the free software OCTAVE, through which we compare the time elapsed for the resolution of a few linear systems, using the factorizations cited. / Mestrado / Algebra linear / Mestre em Matemática

Este trabalho consiste de três partes: Teoremas de coordenatização de Wedderburn e de Zorn, O problema de Nathan Jacobson e Teoremas de Fatorização de Kronecker para as superálgebras alternativas. Na primeira parte apresentamos os teoremas de coordenatização de Wedderburn e de Zorn e suas aplicações na teoria de representações das álgebras associativas e alternativas. Na segunda parte resolvemos um problema de longa data que foi anunciado por Nathan Jacobson sobre a descrição das álgebras alternativas que contém M&#8322(F ) (álgebra associativa de matrizes 2 × 2) com o mesmo elemento identidade. Na terceira parte damos uma prova independente que é válida em qualquer característica do clássico Teorema de Fatorização de Kronecker de Nathan Jacobson. Generalizamos esse resultado e provamos um teorema de Fatorização de Kronecker para as superálgebras alternativas cuja parte par contém O com o mesmo elemento identidade. Além disso, provamos um Teorema de Fatorização de Kronecker para as superálgebras alternativas que contêm a superálgebra associativa M(1|1)(F ) com o mesmo elemento identidade. Como Corolário desse resultado, respondemos a um análogo do problema de Jacobson para as superálgebras alternativas, isto é, descrevemos as superálgebras alternativas que contêm à superálgebra associativa M(1|1)(F ) com o mesmo elemento identidade. Finalmente, estudamos as representações das superálgebras alternativa simples O(4,4) e O[u]. Classificamos os bimodules sobre essas superálgebras e provamos alguns análogos do Teorema de Fatorização de Kronecker para as superálgebras alternativas que contenham O(4|4) ou O[u] com o mesmo elemento identidade / This work consists of three parts: Wedderburn and Zorn coordinatizations theorems, Nathan Jacobsons problem and Kroneckers Factorization theorems for alternative superalgebras. In the first part we present Wedderburn and Zorn coordinatizations theorems and their applications in the theory of representations of associative and alternative algebras. In the second part we solve a long standing problem that was announced by Nathan Jacobson on the description of alternative algebras containing M&#8322(F ) (associative matrix algebra 2 × 2) with the same identity element. In the third part we give an independent proof that is valid in any characteristic of Nathan Jacobsons classic Kronecker Factorization Theorem. We generalize this result and prove a Kronecker Factorization Theorem for alternative superalgebras whose even part contains O with the same identity element. In addition, we prove a Kronecker Factorization Theorem for alternative superalgebras containing the associative superalgebra M(1|1)(F ) with the same identity element. As a corollary of this result, we respond to an analogue of Jacobsons problem for alternative superalgebras, that is, we describe the alternative superalgebras containing the associative superalgebra M(1|1)(F ) with the same identity element. Finally, we study the representations of the simple alternative superalgebras O(4|4) e O[u]. We classify the bimodules on these superalgebras and prove some analogues of the Kronecker Factorization Theorem for alternative superalgebras containing O(4|4) or O[u] with the same identity element

Orientador: Aurelio Ribeiro Leite de Oliveira / Tese (doutorado) - Universidade Estadual de Campinas, Instituto de Matemática Estatística e Computação Científica / Made available in DSpace on 2018-08-25T19:56:24Z (GMT). No. of bitstreams: 1 Silva_LinoMarcosda_D.pdf: 2297954 bytes, checksum: 2213b987c2753edec9152998b30b7c74 (MD5) Previous issue date: 2014 / Resumo: O método de pontos interiores para programação linear resolve em poucas iterações problemas de grande porte. No entanto, requer a cada iteração a resolução de dois sistemas lineares, os quais possuem a mesma matriz de coeficientes. Essa etapa se constitui no passo mais caro do método por aumentar consideravelmente o tempo de processamento e a necessidade de armazenamento de dados. Reduzir o tempo de solução dos sistemas lineares é, portanto, uma forma de melhorar o desempenho do método. De um modo geral, problemas de programação linear de grande porte possuem matrizes esparsas. Uma vez que os sistemas lineares a serem resolvidos são simétricos positivos definidos, métodos iterativos como o método dos gradientes conjugados precondicionado podem ser utilizados na resolução dos mesmos. Além disso, fatores de Cholesky incompletos podem ser utilizados como precondicionadores para o problema. Por outro lado, fatorações incompletas podem sofrer falhas na diagonal durante o processo de fatoração, e quando tais falhas ocorrem uma correção é efetuada somando-se um valor positivo aos elementos da diagonal da matriz do sistema linear e a fatoração da nova matriz é reiniciada, aumentando dessa forma o tempo de precondicionamento, quer seja devido a reconstrução do precondicionador, quer seja devido a perda de qualidade do mesmo. O precondicionador fatoração controlada de Cholesky tem um bom desempenho nas iterações iniciais do método de pontos interiores e tem sido importante nas implementações de abordagens de precondicionamento híbrido. No entanto, sendo uma fatoração incompleta, o mesmo não está livre da ocorrência de falhas no cálculo do pivô. Neste estudo propomos duas modificações à fatoração controlada de Cholesky a fim de evitar ou diminuir o número de reinícios da fatoração das matrizes diagonalmente modificadas. Resultados computacionais mostram que a técnica pode reduzir significativamente o tempo de resolução de certas classes de problemas de programação linear via método de pontos interiores / Abstract: The interior point method solves large linear programming problems in few iterations. However, each iteration requires computing the solution of one or more linear systems. This constitutes the most expensive step of the method by greatly increasing the processing time and the need for data storage. According to it, reducing the time to solve the linear system is a way of improving the method performance. In general, large linear programming problems have sparse matrices. Since the linear systems to be solved are symmetric positive definite, iterative methods such as the preconditioned conjugate gradient method can be used to solve them. Furthermore, incomplete Cholesky factor can be used as a preconditioner to the problem. On the other hand, breakdown may occur during incomplete factorizations. When such failure occur, a correction is made by adding a positive number to diagonal elements of the linear system matrix and the factorization of the new matrix is restarted, thus increasing the time of preconditioning, either due to computing the preconditioner, or due to loss of its quality. The controlled Cholesky factorization preconditioner performs well in early iterations of interior point methods and has been important on implementations of hybrid preconditioning approaches. However, being an incomplete factorization, it is not free from faulty pivots. In this study we propose two modifications to the controlled Cholesky factorization in order to avoid or decrease the refactoring diagonally modified matrices number. Computational results show that the proposed techniques can significantly reduces the time for solving linear programming problems by interior point method / Doutorado / Matematica Aplicada / Doutor em Matemática Aplicada

L’apprentissage par transfert consiste `a utiliser un jeu de taches pour influencerl’apprentissage et améliorer les performances sur une autre tache.Cependant, ce paradigme d’apprentissage peut en réalité gêner les performancessi les taches (sources et cibles) sont trop dissemblables. Un défipour l’apprentissage par transfert est donc de développer des approchesqui détectent et évitent le transfert négatif des connaissances utilisant tr`espeu d’informations sur la tache cible. Un cas particulier de ce type d’apprentissageest l’adaptation de domaine. C’est une situation o`u les tachessources et cibles sont identiques mais dans des domaines différents. Danscette thèse, nous proposons des approches adaptatives basées sur la factorisationmatricielle non-figurative permettant ainsi de trouver une représentationadéquate des données pour ce type d’apprentissage. En effet, unereprésentation utile rend généralement la structure latente dans les donnéesexplicite, et réduit souvent la dimensionnalité´e des données afin que d’autresméthodes de calcul puissent être appliquées. Nos contributions dans cettethèse s’articulent autour de deux dimensions complémentaires : théoriqueet pratique.Tout d’abord, nous avons propose deux méthodes différentes pour résoudrele problème de l’apprentissage par transfert non supervise´e bas´e sur destechniques de factorisation matricielle non-négative. La première méthodeutilise une procédure d’optimisation itérative qui vise `a aligner les matricesde noyaux calculées sur les bases des données provenant de deux taches.La seconde représente une approche linéaire qui tente de découvrir unplongement pour les deux taches minimisant la distance entre les distributionsde probabilité correspondantes, tout en préservant la propriété depositivité.Nous avons également propos´e un cadre théorique bas´e sur les plongementsHilbert-Schmidt. Cela nous permet d’améliorer les résultats théoriquesde l’adaptation au domaine, en introduisant une mesure de distancenaturelle et intuitive avec de fortes garanties de calcul pour son estimation.Les résultats propos´es combinent l’etancheite des bornes de la théoried’apprentissage de Rademacher tout en assurant l’estimation efficace deses facteurs cl´es.Les contributions théoriques et algorithmiques proposées ont et évaluéessur un ensemble de données de référence dans le domaine avec des résultatsprometteurs. / The ability of a human being to extrapolate previously gained knowledge to other domains inspired a new family of methods in machine learning called transfer learning. Transfer learning is often based on the assumption that objects in both target and source domains share some common feature and/or data space. If this assumption is false, most of transfer learning algorithms are likely to fail. In this thesis we propose to investigate the problem of transfer learning from both theoretical and applicational points of view.First, we present two different methods to solve the problem of unsuper-vised transfer learning based on Non-negative matrix factorization tech-niques. First one proceeds using an iterative optimization procedure that aims at aligning the kernel matrices calculated based on the data from two tasks. Second one represents a linear approach that aims at discovering an embedding for two tasks that decreases the distance between the cor-responding probability distributions while preserving the non-negativity property.We also introduce a theoretical framework based on the Hilbert-Schmidt embeddings that allows us to improve the current state-of-the-art theo-retical results on transfer learning by introducing a natural and intuitive distance measure with strong computational guarantees for its estimation. The proposed results combine the tightness of data-dependent bounds de-rived from Rademacher learning theory while ensuring the efficient esti-mation of its key factors.Both theoretical contributions and the proposed methods were evaluated on a benchmark computer vision data set with promising results. Finally, we believe that the research direction chosen in this thesis may have fruit-ful implications in the nearest future.