La langue arabe est pauvre en ressources sémantiques électroniques. Il y a bien la ressource Arabic WordNet, mais il est pauvre en mots et en relations. Cette thèse porte sur l’enrichissement d’Arabic WordNet par des synsets (un synset est un ensemble de mots synonymes) à partir d’un corpus général de grande taille. Ce type de corpus n’existe pas en arabe, il a donc fallu le construire, avant de lui faire subir un certain nombre de prétraitements.Nous avons élaboré, Gilles Bernard et moi-même, une méthode de vectorisation des mots, GraPaVec, qui puisse servir ici. J’ai donc construit un système incluant un module Add2Corpus, des prétraitements, une vectorisation des mots à l’aide de patterns fréquentiels générés automatiquement, qui aboutit à une matrice de données avec en ligne les mots et en colonne les patterns, chaque composante représente la fréquence du mot dans le pattern.Les vecteurs de mots sont soumis au modèle neuronal Self Organizing Map SOM ; la classification produite par SOM construit des synsets. Pour validation, il a fallu créer un corpus de référence (il n’en existe pas en arabe pour ce domaine) à partir d’Arabic WordNet, puis comparer la méthode GraPaVec avec Word2Vec et Glove. Le résultat montre que GraPaVec donne pour ce problème les meilleurs résultats avec une F-mesure supérieure de 25 % aux deux autres. Les classes produites seront utilisées pour créer de nouveaux synsets intégrés à Arabic WordNet / The Arabic language is poor in electronic semantic resources. Among those resources there is Arabic WordNet which is also poor in words and relationships.This thesis focuses on enriching Arabic WordNet by synsets (a synset is a set of synonymous words) taken from a large general corpus. This type of corpus does not exist in Arabic, so we had to build it, before subjecting it to a number of pretreatments.We developed, Gilles Bernard and myself, a method of word vectorization called GraPaVec which can be used here. I built a system which includes a module Add2Corpus, pretreatments, word vectorization using automatically generated frequency patterns, which yields a data matrix whose rows are the words and columns the patterns, each component representing the frequency of a word in a pattern.The word vectors are fed to the neural model Self Organizing Map (SOM) ;the classification produced constructs synsets. In order to validate the method, we had to create a gold standard corpus (there are none in Arabic for this area) from Arabic WordNet, and then compare the GraPaVec method with Word2Vec and Glove ones. The result shows that GraPaVec gives for this problem the best results with a F-measure 25 % higher than the others. The generated classes will be used to create new synsets to be included in Arabic WordNet.

La vidéo sur IP a connu un essor rapide ces dernières années allant de la diffusion télévisuelle en haute qualité via des réseaux dédiés à la diffusion sur internet de contenus vidéo grand public. L’optimisation de l’encodage vidéo H.264/AVC permet aux différents acteurs du marché de se différencier en proposant des solutions pour réduire le débit nécessaire à la représentation d’un flux vidéo ainsi que pour améliorer la qualité perçue par les utilisateurs. C’est dans ce contexte de vidéo professionnelle en haute qualité que s’inscrivent ces travaux de thèse CIFRE réalisés au sein de l’entreprise Digigram, proposant des encodeurs vidéo temps réel pour des diffusions professionnelles en direct. Nous proposons deux solutions de prétraitement pour répondre aux problématiques du secteur de la distribution vidéo. Les deux solutions considèrent les caractéristiques du système visuel humain en exploitant un modèle de JND (Just Noticeable Distortion) définissant des seuils de perception en fonction d’une analyse du contenu des séquences vidéo à encoder. La première solution utilise un préfiltre adaptatif indépendant de l’encodeur, contrôlé par un modèle JND afin d'éliminer le contenu perceptuellement non pertinent et ainsi réduire le débit sans altérer la qualité ressentie. Une analyse approfondie de plusieurs filtres de la littérature, dont le filtre AWA (Adaptive Weighted Averaging) et le filtre bilatéral, nous a également amené à définir deux nouveaux filtres à support étendu qui permettent d’exploiter au mieux les corrélations dans les images haute définition. A l’aide de tests subjectifs, nous montrons que les préfiltres perceptuels proposés permettent en moyenne de diminuer le débit en sortie du codeur d'environ 20% pour une qualité constante en encodage VBR (débit variable) Intra et Inter-image. Finalement, une deuxième solution s’attache à améliorer la qualité perçue dans un contexte d’encodage CBR (débit constant) en intégrant un modèle JND dans l’une des implémentations de la norme H.264/AVC la plus reconnue, le codec x264. Une quantification adaptative perceptuelle est ainsi proposée permettant d’améliorer les performances du codec x264 en améliorant le codage de l’information de contour à moyen et bas débits en encodage intra et inter-image. / The use of digital video over IP has increased exponentially over the last years, due to the development of high-speed networks dedicated to high quality TV transmission as well as the wide development of the nonprofessional video webcast. Optimization of the H.264/AVC encoding process allows manufacturers to offer differentiating encoding solutions, by reducing the bandwidth necessary for transmitting a video sequence at a given quality level, or improving the quality perceived by final users at a fixed bit rate. This thesis was carried out at the company Digigram in a context of professional high quality video. We propose two solutions of preprocessing which consider the characteristics of the human visual system by exploiting a JND profile (Just Noticeable Distortion). A JND model defines perceptual thresholds, below which a distortion cannot be seen, according to the video content. The first solution proposes an adaptive pre-filter independent to the encoder, controlled by a JND profile to reduce the perceptually non-relevant content and so reduce the bitrate while maintaining the perceived quality. By analyzing the state-of-the-art literature, the AWA (Adaptive Weighted Averaging) and Bilateral filters have been selected. Then we define two new filters using a large convolution mask, which enable to better exploit correlations in high-definition video contents. Through subjective tests, we show that the proposed perceptual prefilters give an average bitrate reduction of 20% for the same visual quality in VBR (Variable Bitrate) H.264/AVC Intra and Inter encoding. Finally, the second solution enables to improve the perceived quality in CBR (Constant Bitrate) encoding, by integrating the JND profile into the x264 codec, one of the best implementation of the H.264/AVC standard. Thus, we propose a perceptual adaptive quantization which enhances the x264 performance by improving edge information coding in low and middle bitrate applications.

A técnica de mineração de regras de associação surgiu com o objetivo de encontrar conhecimento novo, útil e previamente desconhecido em bancos de dados transacionais, e uma grande quantidade de algoritmos de mineração de regras de associação tem sido proposta na última década. O maior e mais bem conhecido problema destes algoritmos é a geração de grandes quantidades de conjuntos freqüentes e regras de associação. Em bancos de dados geográficos o problema de mineração de regras de associação espacial aumenta significativamente. Além da grande quantidade de regras e padrões gerados a maioria são associações do domínio geográfico, e são bem conhecidas, normalmente explicitamente representadas no esquema do banco de dados. A maioria dos algoritmos de mineração de regras de associação não garantem a eliminação de dependências geográficas conhecidas a priori. O resultado é que as mesmas associações representadas nos esquemas do banco de dados são extraídas pelos algoritmos de mineração de regras de associação e apresentadas ao usuário. O problema de mineração de regras de associação espacial pode ser dividido em três etapas principais: extração dos relacionamentos espaciais, geração dos conjuntos freqüentes e geração das regras de associação. A primeira etapa é a mais custosa tanto em tempo de processamento quanto pelo esforço requerido do usuário. A segunda e terceira etapas têm sido consideradas o maior problema na mineração de regras de associação em bancos de dados transacionais e tem sido abordadas como dois problemas diferentes: “frequent pattern mining” e “association rule mining”. Dependências geográficas bem conhecidas aparecem nas três etapas do processo. Tendo como objetivo a eliminação dessas dependências na mineração de regras de associação espacial essa tese apresenta um framework com três novos métodos para mineração de regras de associação utilizando restrições semânticas como conhecimento a priori. O primeiro método reduz os dados de entrada do algoritmo, e dependências geográficas são eliminadas parcialmente sem que haja perda de informação. O segundo método elimina combinações de pares de objetos geográficos com dependências durante a geração dos conjuntos freqüentes. O terceiro método é uma nova abordagem para gerar conjuntos freqüentes não redundantes e sem dependências, gerando conjuntos freqüentes máximos. Esse método reduz consideravelmente o número final de conjuntos freqüentes, e como conseqüência, reduz o número de regras de associação espacial. / The association rule mining technique emerged with the objective to find novel, useful, and previously unknown associations from transactional databases, and a large amount of association rule mining algorithms have been proposed in the last decade. Their main drawback, which is a well known problem, is the generation of large amounts of frequent patterns and association rules. In geographic databases the problem of mining spatial association rules increases significantly. Besides the large amount of generated patterns and rules, many patterns are well known geographic domain associations, normally explicitly represented in geographic database schemas. The majority of existing algorithms do not warrant the elimination of all well known geographic dependences. The result is that the same associations represented in geographic database schemas are extracted by spatial association rule mining algorithms and presented to the user. The problem of mining spatial association rules from geographic databases requires at least three main steps: compute spatial relationships, generate frequent patterns, and extract association rules. The first step is the most effort demanding and time consuming task in the rule mining process, but has received little attention in the literature. The second and third steps have been considered the main problem in transactional association rule mining and have been addressed as two different problems: frequent pattern mining and association rule mining. Well known geographic dependences which generate well known patterns may appear in the three main steps of the spatial association rule mining process. Aiming to eliminate well known dependences and generate more interesting patterns, this thesis presents a framework with three main methods for mining frequent geographic patterns using knowledge constraints. Semantic knowledge is used to avoid the generation of patterns that are previously known as non-interesting. The first method reduces the input problem, and all well known dependences that can be eliminated without loosing information are removed in data preprocessing. The second method eliminates combinations of pairs of geographic objects with dependences, during the frequent set generation. A third method presents a new approach to generate non-redundant frequent sets, the maximal generalized frequent sets without dependences. This method reduces the number of frequent patterns very significantly, and by consequence, the number of association rules.

A técnica de mineração de regras de associação surgiu com o objetivo de encontrar conhecimento novo, útil e previamente desconhecido em bancos de dados transacionais, e uma grande quantidade de algoritmos de mineração de regras de associação tem sido proposta na última década. O maior e mais bem conhecido problema destes algoritmos é a geração de grandes quantidades de conjuntos freqüentes e regras de associação. Em bancos de dados geográficos o problema de mineração de regras de associação espacial aumenta significativamente. Além da grande quantidade de regras e padrões gerados a maioria são associações do domínio geográfico, e são bem conhecidas, normalmente explicitamente representadas no esquema do banco de dados. A maioria dos algoritmos de mineração de regras de associação não garantem a eliminação de dependências geográficas conhecidas a priori. O resultado é que as mesmas associações representadas nos esquemas do banco de dados são extraídas pelos algoritmos de mineração de regras de associação e apresentadas ao usuário. O problema de mineração de regras de associação espacial pode ser dividido em três etapas principais: extração dos relacionamentos espaciais, geração dos conjuntos freqüentes e geração das regras de associação. A primeira etapa é a mais custosa tanto em tempo de processamento quanto pelo esforço requerido do usuário. A segunda e terceira etapas têm sido consideradas o maior problema na mineração de regras de associação em bancos de dados transacionais e tem sido abordadas como dois problemas diferentes: “frequent pattern mining” e “association rule mining”. Dependências geográficas bem conhecidas aparecem nas três etapas do processo. Tendo como objetivo a eliminação dessas dependências na mineração de regras de associação espacial essa tese apresenta um framework com três novos métodos para mineração de regras de associação utilizando restrições semânticas como conhecimento a priori. O primeiro método reduz os dados de entrada do algoritmo, e dependências geográficas são eliminadas parcialmente sem que haja perda de informação. O segundo método elimina combinações de pares de objetos geográficos com dependências durante a geração dos conjuntos freqüentes. O terceiro método é uma nova abordagem para gerar conjuntos freqüentes não redundantes e sem dependências, gerando conjuntos freqüentes máximos. Esse método reduz consideravelmente o número final de conjuntos freqüentes, e como conseqüência, reduz o número de regras de associação espacial. / The association rule mining technique emerged with the objective to find novel, useful, and previously unknown associations from transactional databases, and a large amount of association rule mining algorithms have been proposed in the last decade. Their main drawback, which is a well known problem, is the generation of large amounts of frequent patterns and association rules. In geographic databases the problem of mining spatial association rules increases significantly. Besides the large amount of generated patterns and rules, many patterns are well known geographic domain associations, normally explicitly represented in geographic database schemas. The majority of existing algorithms do not warrant the elimination of all well known geographic dependences. The result is that the same associations represented in geographic database schemas are extracted by spatial association rule mining algorithms and presented to the user. The problem of mining spatial association rules from geographic databases requires at least three main steps: compute spatial relationships, generate frequent patterns, and extract association rules. The first step is the most effort demanding and time consuming task in the rule mining process, but has received little attention in the literature. The second and third steps have been considered the main problem in transactional association rule mining and have been addressed as two different problems: frequent pattern mining and association rule mining. Well known geographic dependences which generate well known patterns may appear in the three main steps of the spatial association rule mining process. Aiming to eliminate well known dependences and generate more interesting patterns, this thesis presents a framework with three main methods for mining frequent geographic patterns using knowledge constraints. Semantic knowledge is used to avoid the generation of patterns that are previously known as non-interesting. The first method reduces the input problem, and all well known dependences that can be eliminated without loosing information are removed in data preprocessing. The second method eliminates combinations of pairs of geographic objects with dependences, during the frequent set generation. A third method presents a new approach to generate non-redundant frequent sets, the maximal generalized frequent sets without dependences. This method reduces the number of frequent patterns very significantly, and by consequence, the number of association rules.

Made available in DSpace on 2016-06-02T19:05:46Z (GMT). No. of bitstreams: 1 3250.pdf: 4243437 bytes, checksum: 462620b959061657eacd8e13120fedb0 (MD5) Previous issue date: 2010-09-24 / Financiadora de Estudos e Projetos / There is a large volume of unstructured information (i.e., in text format) being published in electronic media, in digital libraries particularly. Thus, the human being becomes restricted to an amount of text that is able to process and to assimilate over time. In this dissertation is proposed a methodology for textual preprocessing to extract information about disease effects in the biomedical domain papers, in order to identify relevant information from a text, to structure and to store this information in a database to provide a future discovery of interesting relationships between the extracted information. The methodology consists of four steps: Data Entrance (Step 1), Sentence Classification (Step 2), Identification of Relevant Terms (Step 3) and Terms Management (Step 4). This methodology uses three information extraction approaches from the literature: machine learning approach, dictionary-based approach and rule-based approach. The first one is developed in Step 2, in which a supervised machine learning algorithm is responsible for classify the sentences. The second and third ones are developed in Step 3, in which a dictionary of terms validated by an expert and rules developed through regular expressions were used to identify relevant terms in sentences. The methodology validation was carried out through its instantiation to an area of the biomedical domain, more specifically using papers on Sickle Cell Anemia. Accordingly, two case studies were conducted in both Step 2 and in Step 3. The obtained accuracy in the sentence classification was above of 60% and F-measure for the negative effect class was above of 70%. These values correspond to the results achieved with the Support Vector Machine algorithm along with the use of the Noise Removal filter. The obtained F-measure with the identification of relevant terms was above of 85% for the fictitious extraction (i.e., manual classification performed by the expert) and above of 80% for the actual extraction (i.e., automatic classification performed by the classifier). The F-measure of the classifier above of 70% and F-measure of the actual extraction above 80% show the relevance of the sentence classification in the proposed methodology. Importantly to say that many false positives would be identified in full text papers without the sentence classification step. / Existe um grande volume de informação não estruturada (i.e., em formato textual) sendo publicada cada vez mais em meios eletrônicos, particularmente em bibliotecas digitais. Assim, com o passar do tempo, o ser humano fica cada vez mais restringido a uma limitada quantidade de texto que é capaz de processar e assimilar. No sentido de identificar as informações relevantes de um texto e com o objetivo de estruturar e armazenar essas informações em um banco de dados, a fim de propiciar uma futura descoberta de relacionamentos interessantes entre as informações extraídas, nesta dissertação é proposta uma metodologia de pré-processamento textual para extração de informação sobre efeitos de doenças em artigos científicos do domínio biomédico. A metodologia é composta por quatro etapas: Entrada de Dados (Etapa 1), Classificação de Sentenças (Etapa 2), Identificação de Termos Relevantes (Etapa 3) e Gerenciamento de Termos (Etapa 4). Esta metodologia utiliza três abordagens de extração de informação encontradas na literatura: abordagem baseada em aprendizado de máquina, abordagem baseada em dicionário e abordagem baseada em regras. A primeira abordagem é desenvolvida na Etapa 2, na qual um algoritmo de aprendizado de máquina supervisionado é responsável em classificar as sentenças. A segunda e a terceira abordagens são desenvolvidas na Etapa 3, na qual um dicionário de termos validados pelo especialista e regras desenvolvidas por meio de expressões regulares foram utilizados para identificar termos relevantes nas sentenças. A validação da metodologia foi realizada por meio de sua instanciação para uma área do domínio biomédico, mais especificamente usando artigos sobre a doença Anemia Falciforme. Nesse sentido, dois estudos de caso foram realizados tanto na Etapa 2 quanto na Etapa 3. O valor da acurácia obtida na classificação de sentenças foi acima de 60% e o valor da medida-F para a classe efeito negativo foi acima de 70%. Estes valores correspondem aos resultados alcançados com o algoritmo de aprendizado de máquina Support Vector Machine juntamente com a aplicação do filtro Remoção de Ruído. A medida-F obtida com a identificação de termos relevantes foi acima de 85% para a extração fictícia (i.e., classificação manual realizada pelo especialista) e acima de 80% para a extração real (i.e., classificação automática realizada pelo classificador). O valor de medida-F acima de 70% do classificador e o valor de medida-F acima de 80% da extração real mostra a relevância da classificação de sentenças na metodologia proposta. É importante ressaltar que sem a classificação de sentença, muitos falsos positivos seriam identificados nos artigos completos.

Made available in DSpace on 2016-06-02T19:05:56Z (GMT). No. of bitstreams: 1 4310.pdf: 3265738 bytes, checksum: 6650fb70eee9b096860bcac6b5ed596c (MD5) Previous issue date: 2012-02-24 / Currently in the medical field there is a large amount of unstructured information (i.e., in textual format). Regarding the large volume of data, it makes it impossible for doctors and specialists to analyze manually all the relevant literature, which requires techniques for automatically analyze the documents. In order to identify relevant information, as well as to structure and store them into a database and to enable future discovery of significant relationships, in this paper we propose a paragraph-based process to extract treatments from scientific papers in the biomedical domain. The hypothesis is that the initial search for sentences that have terms of complication improves the identification and extraction of terms of treatment. This happens because treatments mainly occur in the same sentence of a complication, or in nearby sentences in the same paragraph. Our methodology employs three approaches for information extraction: machine learning-based approach, for classifying sentences of interest that will have terms to be extracted; dictionary-based approach, which uses terms validated by an expert in the field; and rule-based approach. The methodology was validated as proof of concept, using papers from the biomedical domain, specifically, papers related to Sickle Cell Anemia disease. The proof of concept was performed in the classification of sentences and identification of relevant terms. The value obtained in the classification accuracy of sentences was 79% for the classifier of complication and 71% for the classifier of treatment. These values are consistent with the results obtained from the combination of the machine learning algorithm Support Vector Machine with the filter Noise Removal and Balancing of Classes. In the identification of relevant terms, the results of our methodology showed higher F-measure percentage (42%) compared to the manual classification (31%) and to the partial process, i.e., without using the classifier of complication (36%). Even with low percentage of recall, there was no impact observed on the extraction process, and, in addition, we were able to validate the hypothesis considered in this work. In other words, it was possible to obtain 100% of recall for different terms, thus not impacting the extraction process, and further the working hypothesis of this study was proven. / Atualmente na área médica existe uma grande quantidade de informações não estruturadas (i.e., em formato textual) sendo produzidas na literatura médica. Com o grande volume de dados, torna-se impossível que os médicos e especialistas da área analisem toda a literatura de forma manual, exigindo técnicas para automatizar a análise destes documentos. Com o intuito de identificar as informações relevantes, estruturar e armazenar estas informações em um banco de dados, para posteriormente identificar relacionamentos interessantes entre as informações extraídas, nesta dissertação é proposto um processo baseado em parágrafos para a extração de tratamentos de artigos científicos do domínio biomédico. A hipótese é que a busca inicial de sentenças que possuem termos de complicação melhora a eficiência na identificação e na extração de termos de tratamento. Isso acontece porque tratamentos ocorrem principalmente na mesma sentença de complicação ou em sentenças próximas no mesmo parágrafo. Esta metodologia utiliza três abordagens de extração de informação encontradas na literatura: abordagem baseada em aprendizado de máquina para classificar as sentenças de interesse; abordagem baseada em dicionário com termos validados pelo especialista da área e abordagem baseada em regras. A metodologia foi validada como prova de conceito, utilizando artigos do domínio biomédico, mais especificamente da doença Anemia Falciforme. A prova de conceito foi realizada na classificação de sentenças e identificação de termos relevantes. O valor da acurácia obtida na classificação de sentenças foi de 79% para o classificador de complicação e 71% para o classificador de tratamento. Estes valores condizem com os resultados obtidos com a combinação do algoritmo de aprendizado de máquina Support Vector Machine juntamente com a aplicação do filtro Remoção de Ruído e Balanceamento das Classes. Na identificação de termos relevantes, os resultados da metodologia proposta obteve percentual superior de 42% de medida-F comparado à classificação manual (31%) e comparado ao processo parcial, ou seja, sem utilizar o classificador de complicação (36%). Mesmo com a baixa revocação, foi possível obter 100% de revocação para os termos distintos de tratamento, não impactando o processo de extração, e portanto a hipótese considerada neste trabalho foi comprovada.

Advances in data collection methods, storage and processing technology are providing a unique challenge and opportunity for automated data learning techniques which aim at producing high-level information, or models, from data. A Typical knowledge discovery process consists of data selection, data preparation, data transformation, data mining and interpretation/validation of the results. Thus, we develop automatic learning techniques which contribute to the data preparation, transformation and mining tasks of knowledge discovery. In doing so, we try to improve the prediction accuracy of the overall learning process. Our work focuses on decision tree based learning and thus, we introduce various preprocessing and transformation techniques such as discretization, fuzzy partitioning and dimensionality reduction to improve this type of learning. However, these techniques can be used in other learning methods e.g. discretization can also be used for naive-bayes classifiers. The data preparation step represents almost 80 percent of the problem and is both time consuming and critical for the quality of modeling. Discretization of continuous features is an important problem that has effects on accuracy, complexity, variance and understandability of the induction models. In this thesis, we propose and develop resampling based aggregation techniques that improve the quality of discretization. Later, we validate by comparing with other discretization techniques and with an optimal partitioning method on 10 benchmark data sets.The second part of our thesis concerns with automatic fuzzy partitioning for soft decision tree induction. Soft or fuzzy decision tree is an extension of the classical crisp tree induction such that fuzzy logic is embedded into the induction process with the effect of more accurate models and reduced variance, but still interpretable and autonomous. We modify the above resampling based partitioning method to generate fuzzy partitions. In addition we propose, develop and validate another fuzzy partitioning method that improves the accuracy of the decision tree.Finally, we adopt a topological learning scheme and perform non-linear dimensionality reduction. We modify an existing manifold learning based technique and see whether it can enhance the predictive power and interpretability of classification. / La recherche avancée dans les méthodes d'acquisition de données ainsi que les méthodes de stockage et les technologies d'apprentissage, s'attaquent défi d'automatiser de manière systématique les techniques d'apprentissage de données en vue d'extraire des connaissances valides et utilisables.La procédure de découverte de connaissances s'effectue selon les étapes suivants: la sélection des données, la préparation de ces données, leurs transformation, le fouille de données et finalement l'interprétation et validation des résultats trouvés. Dans ce travail de thèse, nous avons développé des techniques qui contribuent à la préparation et la transformation des données ainsi qu'a des méthodes de fouille des données pour extraire les connaissances. A travers ces travaux, on a essayé d'améliorer l'exactitude de la prédiction durant tout le processus d'apprentissage. Les travaux de cette thèse se basent sur les arbres de décision. On a alors introduit plusieurs approches de prétraitement et des techniques de transformation; comme le discrétisation, le partitionnement flou et la réduction des dimensions afin d'améliorer les performances des arbres de décision. Cependant, ces techniques peuvent être utilisées dans d'autres méthodes d'apprentissage comme la discrétisation qui peut être utilisées pour la classification bayesienne.Dans le processus de fouille de données, la phase de préparation de données occupe généralement 80 percent du temps. En autre, elle est critique pour la qualité de la modélisation. La discrétisation des attributs continus demeure ainsi un problème très important qui affecte la précision, la complexité, la variance et la compréhension des modèles d'induction. Dans cette thèse, nous avons proposes et développé des techniques qui ce basent sur le ré-échantillonnage. Nous avons également étudié d'autres alternatives comme le partitionnement flou pour une induction floue des arbres de décision. Ainsi la logique floue est incorporée dans le processus d'induction pour augmenter la précision des modèles et réduire la variance, en maintenant l'interprétabilité.Finalement, nous adoptons un schéma d'apprentissage topologique qui vise à effectuer une réduction de dimensions non-linéaire. Nous modifions une technique d'apprentissage à base de variété topologiques `manifolds' pour savoir si on peut augmenter la précision et l'interprétabilité de la classification.

A técnica de mineração de regras de associação surgiu com o objetivo de encontrar conhecimento novo, útil e previamente desconhecido em bancos de dados transacionais, e uma grande quantidade de algoritmos de mineração de regras de associação tem sido proposta na última década. O maior e mais bem conhecido problema destes algoritmos é a geração de grandes quantidades de conjuntos freqüentes e regras de associação. Em bancos de dados geográficos o problema de mineração de regras de associação espacial aumenta significativamente. Além da grande quantidade de regras e padrões gerados a maioria são associações do domínio geográfico, e são bem conhecidas, normalmente explicitamente representadas no esquema do banco de dados. A maioria dos algoritmos de mineração de regras de associação não garantem a eliminação de dependências geográficas conhecidas a priori. O resultado é que as mesmas associações representadas nos esquemas do banco de dados são extraídas pelos algoritmos de mineração de regras de associação e apresentadas ao usuário. O problema de mineração de regras de associação espacial pode ser dividido em três etapas principais: extração dos relacionamentos espaciais, geração dos conjuntos freqüentes e geração das regras de associação. A primeira etapa é a mais custosa tanto em tempo de processamento quanto pelo esforço requerido do usuário. A segunda e terceira etapas têm sido consideradas o maior problema na mineração de regras de associação em bancos de dados transacionais e tem sido abordadas como dois problemas diferentes: “frequent pattern mining” e “association rule mining”. Dependências geográficas bem conhecidas aparecem nas três etapas do processo. Tendo como objetivo a eliminação dessas dependências na mineração de regras de associação espacial essa tese apresenta um framework com três novos métodos para mineração de regras de associação utilizando restrições semânticas como conhecimento a priori. O primeiro método reduz os dados de entrada do algoritmo, e dependências geográficas são eliminadas parcialmente sem que haja perda de informação. O segundo método elimina combinações de pares de objetos geográficos com dependências durante a geração dos conjuntos freqüentes. O terceiro método é uma nova abordagem para gerar conjuntos freqüentes não redundantes e sem dependências, gerando conjuntos freqüentes máximos. Esse método reduz consideravelmente o número final de conjuntos freqüentes, e como conseqüência, reduz o número de regras de associação espacial. / The association rule mining technique emerged with the objective to find novel, useful, and previously unknown associations from transactional databases, and a large amount of association rule mining algorithms have been proposed in the last decade. Their main drawback, which is a well known problem, is the generation of large amounts of frequent patterns and association rules. In geographic databases the problem of mining spatial association rules increases significantly. Besides the large amount of generated patterns and rules, many patterns are well known geographic domain associations, normally explicitly represented in geographic database schemas. The majority of existing algorithms do not warrant the elimination of all well known geographic dependences. The result is that the same associations represented in geographic database schemas are extracted by spatial association rule mining algorithms and presented to the user. The problem of mining spatial association rules from geographic databases requires at least three main steps: compute spatial relationships, generate frequent patterns, and extract association rules. The first step is the most effort demanding and time consuming task in the rule mining process, but has received little attention in the literature. The second and third steps have been considered the main problem in transactional association rule mining and have been addressed as two different problems: frequent pattern mining and association rule mining. Well known geographic dependences which generate well known patterns may appear in the three main steps of the spatial association rule mining process. Aiming to eliminate well known dependences and generate more interesting patterns, this thesis presents a framework with three main methods for mining frequent geographic patterns using knowledge constraints. Semantic knowledge is used to avoid the generation of patterns that are previously known as non-interesting. The first method reduces the input problem, and all well known dependences that can be eliminated without loosing information are removed in data preprocessing. The second method eliminates combinations of pairs of geographic objects with dependences, during the frequent set generation. A third method presents a new approach to generate non-redundant frequent sets, the maximal generalized frequent sets without dependences. This method reduces the number of frequent patterns very significantly, and by consequence, the number of association rules.

Knowledge discovery from databases is a complex issue involving integration, data preparation, data mining using machine learning methods and visualization of results. The thesis deals with the whole process of knowledge discovery, especially with the issue of data warehousing, where it offers the design and implementation of a specific data warehouse for the company ROI Hunter, a.s. In the field of data mining, the work focuses on the classification and forecasting of the advertising data available from the prepared data warehouse and, in particular, on the decision tree classification. When predicting the development of new ads, emphasis is put on the rationale for the prediction as well as the proposal to adjust the ad settings so that the prediction ends positively and, with a certain likelihood, the ads actually get better results.

This thesis deals with creation of a mobile library for preprocessing of images with text which represents a part of a system for text recognition. The library is realized with emphasis on generality of use, efficiency and portability. The library providing a set of algorithms primarily for image quality assessment and text detection was created in this thesis. These algorithms enable a substantial decrease in volume of transmitted data and speed up and refinement of the recognition process. An example application for the Android platform able to analyze composition of foods stated on their wrappings was created as well. Overall, the library (system) simplifies development of mobile applications with focus on text extraction and analysis. The mobile application then provides a comfortable way of food harmfulness verification. The thesis offers a reader an overview of current solutions and tools available in this field, it provides a breakdown of important image preprocessing algorithms and guides him through the construction of the library and the application for mobile devices.