Em Aprendizado de Máquina, a abordagem supervisionada normalmente necessita de um número significativo de exemplos de treinamento para a indução de classificadores precisos. Entretanto, a rotulação de dados é freqüentemente realizada manualmente, o que torna esse processo demorado e caro. Por outro lado, exemplos não-rotulados são facilmente obtidos se comparados a exemplos rotulados. Isso é particularmente verdade para tarefas de classificação de textos que envolvem fontes de dados on-line tais como páginas de internet, email e artigos científicos. A classificação de textos tem grande importância dado o grande volume de textos disponível on-line. Aprendizado semi-supervisionado, uma área de pesquisa relativamente nova em Aprendizado de Máquina, representa a junção do aprendizado supervisionado e não-supervisionado, e tem o potencial de reduzir a necessidade de dados rotulados quando somente um pequeno conjunto de exemplos rotulados está disponível. Este trabalho descreve o algoritmo de aprendizado semi-supervisionado co-training, que necessita de duas descrições de cada exemplo. Deve ser observado que as duas descrições necessárias para co-training podem ser facilmente obtidas de documentos textuais por meio de pré-processamento. Neste trabalho, várias extensões do algoritmo co-training foram implementadas. Ainda mais, foi implementado um ambiente computacional para o pré-processamento de textos, denominado PreTexT, com o objetivo de utilizar co-training em problemas de classificação de textos. Os resultados experimentais foram obtidos utilizando três conjuntos de dados. Dois conjuntos de dados estão relacionados com classificação de textos e o outro com classificação de páginas de internet. Os resultados, que variam de excelentes a ruins, mostram que co-training, similarmente a outros algoritmos de aprendizado semi-supervisionado, é afetado de maneira bastante complexa pelos diferentes aspectos na indução dos modelos. / In Machine Learning, the supervised approach usually requires a large number of labeled training examples to learn accurately. However, labeling is often manually performed, making this process costly and time-consuming. By contrast, unlabeled examples are often inexpensive and easier to obtain than labeled examples. This is especially true for text classification tasks involving on-line data sources, such as web pages, email and scientific papers. Text classification is of great practical importance today given the massive volume of online text available. Semi-supervised learning, a relatively new area in Machine Learning, represents a blend of supervised and unsupervised learning, and has the potential of reducing the need of expensive labeled data whenever only a small set of labeled examples is available. This work describes the semi-supervised learning algorithm co-training, which requires a partitioned description of each example into two distinct views. It should be observed that the two different views required by co-training can be easily obtained from textual documents through pre-processing. In this works, several extensions of co-training algorithm have been implemented. Furthermore, we have also implemented a computational environment for text pre-processing, called PreTexT, in order to apply the co-training algorithm to text classification problems. Experimental results using co-training on three data sets are described. Two data sets are related to text classification and the other one to web-page classification. Results, which range from excellent to poor, show that co-training, similarly to other semi-supervised learning algorithms, is affected by modelling assumptions in a rather complicated way.

Les travaux de cette thèse ont été initiés par des problèmes d’apprentissage de données radiomiques. La Radiomique est une discipline médicale qui vise l’analyse à grande échelle de données issues d’imageries médicales traditionnelles, pour aider au diagnostic et au traitement des cancers. L’hypothèse principale de cette discipline est qu’en extrayant une grande quantité d’informations des images, on peut caractériser de bien meilleure façon que l’œil humain les spécificités de cette pathologie. Pour y parvenir, les données radiomiques sont généralement constituées de plusieurs types d’images et/ou de plusieurs types de caractéristiques (images, cliniques, génomiques). Cette thèse aborde ce problème sous l’angle de l’apprentissage automatique et a pour objectif de proposer une solution générique, adaptée à tous problèmes d’apprentissage du même type. Nous identifions ainsi en Radiomique deux problématiques d’apprentissage: (i) l’apprentissage de données en grande dimension et avec peu d’instances (high dimension, low sample size, a.k.a.HDLSS) et (ii) l’apprentissage multi-vues. Les solutions proposées dans ce manuscrit exploitent des représentations de dissimilarités obtenues à l’aide des Forêts Aléatoires. L’utilisation d’une représentation par dissimilarité permet de contourner les difficultés inhérentes à l’apprentissage en grande dimension et facilite l’analyse conjointe des descriptions multiples (les vues). Les contributions de cette thèse portent sur l’utilisation de la mesure de dissimilarité embarquée dans les méthodes de Forêts Aléatoires pour l’apprentissage multi-vue de données HDLSS. En particulier, nous présentons trois résultats: (i) la démonstration et l’analyse de l’efficacité de cette mesure pour l’apprentissage multi-vue de données HDLSS; (ii) une nouvelle méthode pour mesurer les dissimilarités à partir de Forêts Aléatoires, plus adaptée à ce type de problème d’apprentissage; et (iii) une nouvelle façon d’exploiter l’hétérogénèité des vues, à l’aide d’un mécanisme de combinaison dynamique. Ces résultats ont été obtenus sur des données radiomiques mais aussi sur des problèmes multi-vue classiques. / The work of this thesis was initiated by a Radiomic learning problem. Radiomics is a medical discipline that aims at the large-scale analysis of data from traditional medical imaging to assist in the diagnosis and treatment of cancer. The main hypothesis of this discipline is that by extracting a large amount of information from the images, we can characterize the specificities of this pathology in a much better way than the human eye. To achieve this, Radiomics data are generally based on several types of images and/or several types of features (from images, clinical, genomic). This thesis approaches this problem from the perspective of Machine Learning (ML) and aims to propose a generic solution, adapted to any similar learning problem. To do this, we identify two types of ML problems behind Radiomics: (i) learning from high dimension, low sample size (HDLSS) and (ii) multiview learning. The solutions proposed in this manuscript exploit dissimilarity representations obtained using the Random Forest method. The use of dissimilarity representations makes it possible to overcome the well-known difficulties of learning high dimensional data, and to facilitate the joint analysis of the multiple descriptions, i.e. the views.The contributions of this thesis focus on the use of the dissimilarity easurement embedded in the Random Forest method for HDLSS multi-view learning. In particular, we present three main results: (i) the demonstration and analysis of the effectiveness of this measure for HDLSS multi-view learning; (ii) a new method for measuring dissimilarities from Random Forests, better adapted to this type of learning problem; and (iii) a new way to exploit the heterogeneity of views, using a dynamic combination mechanism. These results have been obtained on radiomic data but also on classical multi-view learning problems.