Cette dissertation présente un ensemble d'algorithmes visant à en permettre un usage rapide, robuste et automatique des « Support Vector Machines » (SVM) non supervisés dans un contexte d'analyse de données. Les SVM non supervisés se déclinent sous deux types algorithmes prometteurs, le « Support Vector Clustering » (SVC) et le « Support Vector Domain Description » (SVDD), offrant respectivement une solution à deux problèmes importants en analyse de données, soit la recherche de groupements homogènes (« clustering »), ainsi que la reconnaissance d'éléments atypiques (« novelty/abnomaly detection ») à partir d'un ensemble de données. Cette recherche propose des solutions concrètes à trois limitations fondamentales inhérentes à ces deux algorithmes, notamment I) l'absence d'algorithme d'optimisation efficace permettant d'exécuter la phase d'entrainement des SVDD et SVC sur des ensembles de données volumineux dans un délai acceptable, 2) le manque d'efficacité et de robustesse des algorithmes existants de partitionnement des données pour SVC, ainsi que 3) l'absence de stratégies de sélection automatique des hyperparamètres pour SVDD et SVC contrôlant la complexité et la tolérance au bruit des modèles générés. La résolution individuelle des trois limitations mentionnées précédemment constitue les trois axes principaux de cette thèse doctorale, chacun faisant l'objet d'un article scientifique proposant des stratégies et algorithmes permettant un usage rapide, robuste et exempt de paramètres d'entrée des SVDD et SVC sur des ensembles de données arbitraires.

No description available.

The study explores binary classification with Support Vector Machines as means to predict a satisfaction score based on customer surveys in the customer supportdomain. Standard feature selection methods and their impact on results are evaluated and a feature scoring metric Log Odds Ratio is implemented for addressingasymmetrical class distributions. Results show that the feature selection andscoring methods implemented improve performance significantly. Results alsoshow that it is possible to get decent predictive values on test data based onlimited amount of training observations. However mixed results are presentedin a real-world application example as a there is a significant error rate fordiscriminating the minority class. We also show the negative effects of usingcommon metrics such as accuracy and f-measure for optimizing models whendealing with high-skew data in a classification context.

Computer-based drug design is a vital area of pharmaceutical chemistry; Quantitative Structure-Activity Relationships (QSARs), determined computationally from experimental observations, are crucial in identifying candidate drugs by early screening, saving time on synthesis and in vivo testing. This thesis investigates the viability and the practicalities of using Mass Spectra-based pseudo-molecular descriptors, in comparison with other molecular descriptor systems, to predict the carcinogenicity, mutagenicity and the Cltransport inhibiting ability of a variety of molecules, and in the first case, of chemotherapeutic drugs particularly. It does so by identifying a number of QSARs which link the physical properties of chemicals with their concomitant activities in a reliable and mathematical manner. First, this thesis confirms that carcinogenicity and mutagenicity are indeed predictable using a variety of Artificial Intelligence techniques, both supervised and unsupervised, information germane to pharmaceutical research groups interested in the preliminary screening of candidate anti-cancer drugs. Secondly, it demonstrates that Mass Spectral intensities possess great descriptive fidelity and shows that reducing the burden of dimensionality is not only important, but imperative; selecting this smaller set of orthogonal descriptors is best achieved using Principal Component Analysis as opposed to the selection of a set of the most frequent fragments, or the use of every peak up to a number determined by the boundaries of supervised learning. Thirdly, it introduces a novel system of backpropagation and demonstrates that it is more efficient than its principal competitor at monitoring a series of connection weights when applied to this area of research, which requires complex relationships. Finally, it promulgates some preliminary conclusions about which AI techniques are applicable to certain problem-scenarios, how these techniques might be applied, and the likelihood that that application will result in the identification of series of reliable QSARs.

Sentiment analysis has emerged as a field that has attracted a significant amount of attention since it has a wide variety of applications that could benefit from its results, such as news analytics, marketing, question answering, knowledge management and so on. This area, however, is still early in its development where urgent improvements are required on many issues, particularly on the performance of sentiment classification. In this thesis, three key challenging issues affecting sentiment classification are outlined and innovative ways of addressing these issues are presented. First, text pre-processing has been found crucial on the sentiment classification performance. Consequently, a combination of several existing preprocessing methods is proposed for the sentiment classification process. Second, text properties of financial news are utilised to build models to predict sentiment. Two different models are proposed, one that uses financial events to predict financial news sentiment, and the other uses a new interesting perspective that considers the opinion reader view, as opposed to the classic approach that examines the opinion holder view. A new method to capture the reader sentiment is suggested. Third, one characteristic of financial news is that it stretches over a number of domains, and it is very challenging to infer sentiment between different domains. Various approaches for cross-domain sentiment analysis have been proposed and critically evaluated.

A dissertation submitted in ful llment of the requirements for the degree of Master of Science in the School of Computer Science and Applied Mathematics Faculty of Science August 15, 2017 / Facial expression recognition is one of the most common means of communication used for complementing spoken word. However, people have grown to master ways of ex- hibiting deceptive expressions. Hence, it is imperative to understand di erences in expressions mostly for security purposes among others. Traditional methods employ machine learning techniques in di erentiating real and fake expressions. However, this approach does not always work as human subjects can easily mimic real expressions with a bit of practice. This study presents an approach that evaluates the time related dis- tance that exists between eye actions and an exhibited expression. The approach gives insights on some of the most fundamental characteristics of expressions. The study fo- cuses on nding and understanding the temporal relationship that exists between eye blinks and smiles. It further looks at the relationship that exits between eye closure and pain expressions. The study incorporates active appearance models (AAM) for feature extraction and support vector machines (SVM) for classi cation. It tests extreme learn- ing machines (ELM) in both smile and pain studies, which in turn, attains excellent results than predominant algorithms like the SVM. The study shows that eye blinks are highly correlated with the beginning of a smile in posed smiles while eye blinks are highly correlated with the end of a smile in spontaneous smiles. A high correlation is observed between eye closure and pain in spontaneous pain expressions. Furthermore, this study brings about ideas that lead to potential applications such as lie detection systems, robust health care monitoring systems and enhanced animation design systems among others. / MT 2018

Made available in DSpace on 2014-06-12T17:39:31Z (GMT). No. of bitstreams: 2 arquivo519_1.pdf: 8081494 bytes, checksum: ba2427869f3a7c683fd296629937e553 (MD5) license.txt: 1748 bytes, checksum: 8a4605be74aa9ea9d79846c1fba20a33 (MD5) Previous issue date: 2010 / Coordenação de Aperfeiçoamento de Pessoal de Nível Superior / Um projeto de inovação tecnológica consiste em uma série de análises e procedimentos que tem como fim estimar o valor de uma tecnologia, ou seja, gerar uma estimativa dos rendimentos futuros que tal empreendimento/tecnologia possa proporcionar. A abordagem tradicional da análise de investimentos para esta categoria de projetos possui uma limitação no que tange a: 1 estimação do valor da tecnologia que exige a incorporação de variáveis qualitativas que não são consideradas por essa modelagem; e 2 a elevada variabilidade das estimativas do fluxo de caixa projetado, em virtude das diferentes categorias de risco inerentes a esse tipo de projeto. A partir desta limitação apresentada no estado da arte da avaliação deste tipo de projeto, esta pesquisa de cunho exploratório pretende utilizar a metodologia de rating como uma alternativa a avaliação de projetos de inovação. Pois um sistema de classificação através de rating possui a flexibilidade necessária para a incorporação de variáveis qualitativas que podem auxiliar na mensuração do valor da tecnologia, bem como fornece uma série de procedimentos que permitem a estimação do risco de tais projetos. Tal aplicação da metodologia de rating gera o Sistema de Classificação de Risco de Projetos de Inovação (SCRP), que a partir de uma amostra de 40 projetos de investimento industrial fornecidos pelo Banco do Nordeste do Brasil, indicadores setoriais, macroeconômicos e tecnológicos, provê uma classificação de viabilidade e risco para tais projetos. As Support Vector Machines, técnica de inteligência artificial com resultados exitosos em várias áreas das finanças, inclusive com ratings é introduzida nesta pesquisa para testar a classificação gerada pelo SCRP. A aplicação do SVM fez uso do código LIBSVM e do Software Matlab. A classificação obtida pelo SCRP apresentou um ajuste médio de 83,6% quando comparado aos 10 melhores projetos classificados pelo critério da TIR e de 87,6% de ajuste médio para com os 8 piores projetos classificados pelo critério do VPL, a classificação obtida através do SVM, apresentou uma acuracia de 37,5% frente aos dados de teste

Navigating through the debris of the information explosion requires powerful, flexible search tools. These tools must be both useful and useable; that is, they must do their jobs effectively without placing too many burdens on the user. While general interest search engines, such as Google, have addressed this latter challenge well, more topic-specific search engines, such as PubMed, have not. These search engines, though effective, often require training in their use, as well as in-depth knowledge of the domain over which they operate. Furthermore, search results are often returned in an order irrespective of users' preferences, forcing them to manually search through search results in order to find the documents they find most useful. To solve these problems, we intend to learn ranking functions from user relevance preferences. Applying these ranking functions to search results allows us to improve search usability without having to reengineer existing, effective search engines. Using ranking SVMs and active learning techniques, we can effectively learn what is relevant to a user from relatively small amounts of preference data, and apply these learned models as ranking functions. This gives users the convenience of seeing relevance-ordered search results, which are tailored to their preferences as opposed to using a one-size-fits-all sorting method. As giving preference feedback does not require in-depth domain knowledge, this approach is suitable for use by domain experts as well as neophytes. Furthermore, giving preference feedback does not require a great deal of training, adding very little overhead to the search process.

We address the problem of integrating objects from a source taxonomy into a master taxonomy. This problem is not only pervasive on the nowadays web, but also important to the emerging semantic web. A straightforward approach to automating this process would be to train a classifier for each category in the master taxonomy, and then classify objects from the source taxonomy into these categories. In this paper we attempt to use a powerful classification method, Support Vector Machine (SVM), to attack this problem. Our key insight is that the availability of the source taxonomy data could be helpful to build better classifiers in this scenario, therefore it would be beneficial to do transductive learning rather than inductive learning, i.e., learning to optimize classification performance on a particular set of test examples. Noticing that the categorization of the master and source taxonomies often have some semantic overlap, we propose a new method, Cluster Shrinkage (CS), to further enhance the classification by exploiting such implicit knowledge. Our experiments with real-world web data show substantial improvements in the performance of taxonomy integration. / Singapore-MIT Alliance (SMA)

<p> <p>&nbsp / </p> </p> <p align="left">IP multicast is an efficient mechanism for simultaneously transmitting bulk data to multiple receivers. Many applications can benefit from multicast, such as audio and videoconferencing, multi-player games, multimedia broadcasting, distance education, and data replication. For either technical or policy reasons, IP multicast still has not yet been deployed in today&rsquo / s Internet. Congestion is one of the most important issues impeding the development and deployment of IP multicast and multicast applications.</p>