De nombreux jeux de données sont publiés sur le web à l’aide des technologies du web sémantique. Ces jeux de données contiennent des données qui représentent des liens vers des ressources similaires. Si ces jeux de données sont liés entre eux par des liens construits correctement, les utilisateurs peuvent facilement interroger des données à travers une interface uniforme, comme s’ils interrogeaient un jeu de données unique. Mais, trouver des liens corrects est très difficile car de nombreuses comparaisons doivent être effectuées. Plusieurs solutions ont été proposées pour résoudre ce problème : (1) l’approche la plus directe est de comparer les valeurs d’attributs d’instances pour identifier les liens, mais il est impossible de comparer toutes les paires possibles de valeurs d’attributs. (2) Une autre stratégie courante consiste à comparer les instances selon les attribut correspondants trouvés par l’alignement d’ontologies à base d’instances, qui permet de générer des correspondances d’attributs basés sur des instances. Cependant, il est difficile d’identifier des instances similaires à travers les ensembles de données car,dans certains cas, les valeurs des attributs en correspondance ne sont pas les mêmes.(3) Plusieurs méthodes utilisent la programmation génétique pour construire des modèles d’interconnexion afin de comparer différentes instances, mais elles souffrent de longues durées d’exécution.Dans cette thèse, une méthode d’interconnexion est proposée pour relier les instances similaires dans différents ensembles de données, basée à la fois sur l’apprentissage statistique et sur l’apprentissage symbolique. L’entrée est constituée de deux ensembles de données, des correspondances de classes sur les deux ensembles de données et un échantillon de liens “positif” ou “négatif” résultant d’une évaluation de l’utilisateur. La méthode construit un classifieur qui distingue les bons liens des liens incorrects dans deux ensembles de données RDF en utilisant l’ensemble des liens d’échantillons évalués. Le classifieur est composé de correspondances d’attributs entre les classes correspondantes et de deux ensembles de données,qui aident à comparer les instances et à établir les liens. Le classifieur est appelé motif d’interconnexion dans cette thèse. D’une part, notre méthode découvre des correspondances potentielles entre d’attributs pour chaque correspondance de classe via une méthode d’apprentissage statistique : l’algorithme de regroupement K-medoids,en utilisant des statistiques sur les valeurs des instances. D’autre part, notre solution s’appuie sur un modèle d’interconnexion par une méthode d’apprentissage symbolique: l’espace des versions, basée sur les correspondances d’attributs potentielles découvertes et l’ensemble des liens de l’échantillon évalué. Notre méthode peut résoudre la tâche d’interconnexion quand il n’existe pas de motif d’interconnexion combiné qui couvre tous les liens corrects évalués avec un format concis.L’expérimentation montre que notre méthode d’interconnexion, avec seulement1% des liens totaux dans l’échantillon, atteint une F-mesure élevée (de 0,94 à 0,99). / There are many data sets being published on the web with Semantic Web technology. The data sets usually contain analogous data which represent the similar resources in the world. If these data sets are linked together by correctly identifying the similar instances, users can conveniently query data through a uniform interface, as if they are connecting a single database. However, finding correct links is very challenging because web data sources usually have heterogeneous ontologies maintained by different organizations. Many existing solutions have been proposed for this problem. (1) One straight-forward idea is to compare the attribute values of instances for identifying links, yet it is impossible to compare all possible pairs of attribute values. (2) Another common strategy is to compare instances with correspondences found by instance-based ontology matching, which can generate attribute correspondences based on overlapping ranges between two attributes, while it is easy to cause incomparable attribute correspondences or undiscovered comparable attribute correspondences. (3) Many existing solutions leverage Genetic Programming to construct interlinking patterns for comparing instances, however the running times of the interlinking methods are usually long. In this thesis, an interlinking method is proposed to interlink instances for different data sets, based on both statistical learning and symbolic learning. On the one hand, the method discovers potential comparable attribute correspondences of each class correspondence via a K-medoids clustering algorithm with instance value statistics. We adopt K-medoids because of its high working efficiency and high tolerance on irregular data and even incorrect data. The K-medoids classifies attributes of each class into several groups according to their statistical value features. Groups from different classes are mapped when they have similar statistical value features, to determine potential comparable attribute correspondences. The clustering procedure effectively narrows the range of candidate attribute correspondences. On the other hand, our solution also leverages a symbolic learning method, called Version Space. Version Space is an iterative learning model that searches for the interlinking pattern from two directions. Our design can solve the interlinking task that does not have a single compatible conjunctive interlinking pattern that covers all assessed correct links with a concise format. The interlinking solution is evaluated with large-scale real-world data from IM@OAEI and CKAN. Experiments confirm that the solution with only 1% of sample links already reaches a high accuracy (up to 0.94-0.99 on F-measure). The F-measure quickly converges improving on other state-of-the-art approaches, by nearly 10 percent of their F-measure values.

The purpose of this thesis has been to evaluate if a new instance based explanation method, called Automatic Instance Text Classification Explanator (AITCE), could provide researchers with insights about the predictions of automatic text classification and decision support about documents requiring human classification. Making it possible for researchers, that normally use manual classification, to cut time and money in their research, with the maintained quality. In the study, AITCE was implemented and applied to the predictions of a black box classifier. The evaluation was performed at two levels: at instance level, where a group of 3 senior researchers, that use human classification in their research, evaluated the results from AITCE from an expert view; and at model level, where a group of 24 non experts evaluated the characteristics of the classes. The evaluations indicate that AITCE produces insights about which words that most strongly affect the prediction. The research also suggests that the quality of an automatic text classification may increase through an interaction between the user and the classifier in situations with unsure predictions.

The aim of this research project was to provide a bio-sensing component for a real-time adaptive technology in the context of cultural heritage. The proposed system was designed to infer the interest or intention of the user and to augment elements of the cultural heritage experience interactively through implicit interaction. Implicit interaction in this context is the process whereby the system observes the user while they interact with artefacts; recording psychophysiological responses to cultural heritage artefacts or materials and acting upon these responses to drive adaptations in content in real-time. Real-time biocybernetic control is the central component of physiological computing wherein physiological data are converted into a control input for a technological system. At its core the bio-sensing component is a biocybernetic control loop that utilises an inference of user interest as its primary driver. A biocybernetic loop is composed of four main stages: inference, classification, adaptation and interaction. The programme of research described in this thesis is concerned primarily with exploration of the inference and classification elements of the biocybernetic loop but also encompasses an element of adaptation and interaction. These elements are explored first through literature review and discussion (presented in chapters 1-5) and then through experimental studies (presented in chapters 7-11).

Submitted by ADRIANO BRESSANE null (adrianobressane@ymail.com) on 2017-04-06T11:45:45Z No. of bitstreams: 1 3.pdf: 38954890 bytes, checksum: acc45aa06079de5294c6da5f275e4318 (MD5) / Approved for entry into archive by Luiz Galeffi (luizgaleffi@gmail.com) on 2017-04-12T18:48:05Z (GMT) No. of bitstreams: 1 bressane_a_dr_soro.pdf: 38954890 bytes, checksum: acc45aa06079de5294c6da5f275e4318 (MD5) / Made available in DSpace on 2017-04-12T18:48:05Z (GMT). No. of bitstreams: 1 bressane_a_dr_soro.pdf: 38954890 bytes, checksum: acc45aa06079de5294c6da5f275e4318 (MD5) Previous issue date: 2017-03-31 / Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES) / Embora fundamental para diversas finalidades, a identificação de espécies arbóreas pode ser complexa e até mesmo inviável em determinadas condições, motivando o desenvolvimento de métodos assistidos por inteligência computacional. Nesse sentido, estudos têm se concentrado na avaliação de características extraídas a partir de imagens da folha e, apesar dos avanços, não são aplicáveis a espécies caducifólias em determinadas épocas do ano. Logo, o uso de características baseadas na textura em imagens do tronco poderia ser uma alternativa, mas ainda há poucos resultados reportados na literatura. Portanto, a partir da revisão de trabalhos anteriores, foram realizados experimentos para avaliar o uso de métodos de inteligência computacional no reconhecimento de padrões de textura em imagens do tronco arbóreo. Para tanto, foram consideradas espécies arbóreas caducifólias nativas da flora brasileira. As primeiras análises experimentais focaram na avaliação de padrões. Como resultado, verificou-se que a melhor capacidade de generalização é alcançada combinando o uso de estatísticas de primeira e segunda ordem. Contudo, o aumento de variáveis preditoras demandou uma abordagem capaz de lidar com informação redundante. Entre as técnicas avaliadas para essa finalidade, a análise fatorial exploratória proporcionou redução na taxa de erros durante o aprendizado de máquina e aumento da acurácia durante a validação com dados de teste. Por fim, constatando que a variabilidade natural da textura no tronco arbóreo causa uma ambiguidade no reconhecimento de padrões, o uso da modelagem fuzzy foi avaliado. Em comparação com outros algoritmos de aprendizagem de máquina, a abordagem fuzzy proporcionou resultados competitivos e, assim, pode ser considerada uma alternativa promissora para novos avanços no apoio a identificação de espécies arbóreas usando inteligência computacional. / Although the arboreal identification is mandatory for several purposes, it can be complex and infeasible under certain conditions, motivating the development of computer-aided methods. In this sense, studies have focused on the assessment of features extracted from leaf images and, despite advancements, they are not applicable for deciduous species in some periods of year. Therefore, the usage of features based on texture in trunk images could be an alternative, but there are still few outcomes reported in the literature. Thus, from the review on previous studies, experiments have been performed for evaluating the use of computational intelligence methods for texture patterns recognition in trunk images. For that, native species from the deciduous Brazilian forest were considered. Firstly, the experimental analyzes focused on the evaluation of patterns. As a result, it was noted that the best generalization ability is reached using the first-order statistics in combination with second-order descriptors. Nevertheless, the increase of predictor variables required an approach capable of dealing with redundant information. Among the techniques assessed for this purpose, the exploratory factor analysis provided an error rate reduction during the machine learning, and an accuracy improvement in the validation over testing dataset. Finally, taking into account that the natural variability of texture in arboreal trunk causes an ambiguity in the pattern recognition, the usage of fuzzy modeling has been evaluated. In comparison with other machine learning algorithms, the fuzzy approach afforded competitive results, and hence it can be a promising alternative for further progress in the arboreal identification supported by computational intelligence.

abstract: Reinforcement learning (RL) is a powerful methodology for teaching autonomous agents complex behaviors and skills. A critical component in most RL algorithms is the reward function -- a mathematical function that provides numerical estimates for desirable and undesirable states. Typically, the reward function must be hand-designed by a human expert and, as a result, the scope of a robot's autonomy and ability to safely explore and learn in new and unforeseen environments is constrained by the specifics of the designed reward function. In this thesis, I design and implement a stateful collision anticipation model with powerful predictive capability based upon my research of sequential data modeling and modern recurrent neural networks. I also develop deep reinforcement learning methods whose rewards are generated by self-supervised training and intrinsic signals. The main objective is to work towards the development of resilient robots that can learn to anticipate and avoid damaging interactions by combining visual and proprioceptive cues from internal sensors. The introduced solutions are inspired by pain pathways in humans and animals, because such pathways are known to guide decision-making processes and promote self-preservation. A new "robot dodge ball' benchmark is introduced in order to test the validity of the developed algorithms in dynamic environments. / Dissertation/Thesis / Masters Thesis Computer Science 2018

Gamification is an area that has attracted a substantial amount of interest in recent years and is something that can be applied almost everywhere. Health is also an area that is frequently discussed regarding how a person's health and daily exercise can be improved. Applying gamification to motivate people to exercise more is something that has been done before although they are usually marketed as health application. What if an application's main purpose was not to count the amount calories burnt but instead the exercise is an additional part? In this thesis you will find the development of a prototype that is a gamified Android application with the purpose to motivate people to walk more in stairs. In order to develop an application there was a need for a stairstep counter to count the number of steps a person has taken in a stair. Hence a stairstep counter was developed using the open source machine learning framework, TensorFlow, with classification and neural networks. The counter uses the y- and z-axis of the mobile accelerometer to classify if a movement of a person was a step or not. It was trained with a custom dataset that was created using data from walking in a stair and data collected from other movements. Resulting stairstep counter used in the application is a counter that counts stairsteps rather accurately. The final result of this thesis was a mobile application that uses animation and competition elements to enhance a person's stair walk experience.

abstract: Drosophila melanogaster, as an important model organism, is used to explore the mechanism which governs cell differentiation and embryonic development. Understanding the mechanism will help to reveal the effects of genes on other species or even human beings. Currently, digital camera techniques make high quality Drosophila gene expression imaging possible. On the other hand, due to the advances in biology, gene expression images which can reveal spatiotemporal patterns are generated in a high-throughput pace. Thus, an automated and efficient system that can analyze gene expression will become a necessary tool for investigating the gene functions, interactions and developmental processes. One investigation method is to compare the expression patterns of different developmental stages. Recently, however, the expression patterns are manually annotated with rough stage ranges. The work of annotation requires professional knowledge from experienced biologists. Hence, how to transfer the domain knowledge in biology into an automated system which can automatically annotate the patterns provides a challenging problem for computer scientists. In this thesis, the problem of stage annotation for Drosophila embryo is modeled in the machine learning framework. Three sparse learning algorithms and one ensemble algorithm are used to attack the problem. The sparse algorithms are Lasso, group Lasso and sparse group Lasso. The ensemble algorithm is based on a voting method. Besides that the proposed algorithms can annotate the patterns to stages instead of stage ranges with high accuracy; the decimal stage annotation algorithm presents a novel way to annotate the patterns to decimal stages. In addition, some analysis on the algorithm performance are made and corresponding explanations are given. Finally, with the proposed system, all the lateral view BDGP and FlyFish images are annotated and several interesting applications of decimal stage value are revealed. / Dissertation/Thesis / M.S. Computer Science 2012

abstract: Learning from high dimensional biomedical data attracts lots of attention recently. High dimensional biomedical data often suffer from the curse of dimensionality and have imbalanced class distributions. Both of these features of biomedical data, high dimensionality and imbalanced class distributions, are challenging for traditional machine learning methods and may affect the model performance. In this thesis, I focus on developing learning methods for the high-dimensional imbalanced biomedical data. In the first part, a sparse canonical correlation analysis (CCA) method is presented. The penalty terms is used to control the sparsity of the projection matrices of CCA. The sparse CCA method is then applied to find patterns among biomedical data sets and labels, or to find patterns among different data sources. In the second part, I discuss several learning problems for imbalanced biomedical data. Note that traditional learning systems are often biased when the biomedical data are imbalanced. Therefore, traditional evaluations such as accuracy may be inappropriate for such cases. I then discuss several alternative evaluation criteria to evaluate the learning performance. For imbalanced binary classification problems, I use the undersampling based classifiers ensemble (UEM) strategy to obtain accurate models for both classes of samples. A small sphere and large margin (SSLM) approach is also presented to detect rare abnormal samples from a large number of subjects. In addition, I apply multiple feature selection and clustering methods to deal with high-dimensional data and data with highly correlated features. Experiments on high-dimensional imbalanced biomedical data are presented which illustrate the effectiveness and efficiency of my methods. / Dissertation/Thesis / M.S. Computer Science 2013

abstract: Multidimensional data have various representations. Thanks to their simplicity in modeling multidimensional data and the availability of various mathematical tools (such as tensor decompositions) that support multi-aspect analysis of such data, tensors are increasingly being used in many application domains including scientific data management, sensor data management, and social network data analysis. Relational model, on the other hand, enables semantic manipulation of data using relational operators, such as projection, selection, Cartesian-product, and set operators. For many multidimensional data applications, tensor operations as well as relational operations need to be supported throughout the data life cycle. In this thesis, we introduce a tensor-based relational data model (TRM), which enables both tensor- based data analysis and relational manipulations of multidimensional data, and define tensor-relational operations on this model. Then we introduce a tensor-relational data management system, so called, TensorDB. TensorDB is based on TRM, which brings together relational algebraic operations (for data manipulation and integration) and tensor algebraic operations (for data analysis). We develop optimization strategies for tensor-relational operations in both in-memory and in-database TensorDB. The goal of the TRM and TensorDB is to serve as a single environment that supports the entire life cycle of data; that is, data can be manipulated, integrated, processed, and analyzed. / Dissertation/Thesis / Doctoral Dissertation Computer Science 2014

abstract: Research in the learning sciences suggests that students learn better by collaborating with their peers than learning individually. Students working together as a group tend to generate new ideas more frequently and exhibit a higher level of reasoning. In this internet age with the advent of massive open online courses (MOOCs), students across the world are able to access and learn material remotely. This creates a need for tools that support distant or remote collaboration. In order to build such tools we need to understand the basic elements of remote collaboration and how it differs from traditional face-to-face collaboration. The main goal of this thesis is to explore how spoken dialogue varies in face-to-face and remote collaborative learning settings. Speech data is collected from student participants solving mathematical problems collaboratively on a tablet. Spoken dialogue is analyzed based on conversational and acoustic features in both the settings. Looking for collaborative differences of transactivity and dialogue initiative, both settings are compared in detail using machine learning classification techniques based on acoustic and prosodic features of speech. Transactivity is defined as a joint construction of knowledge by peers. The main contributions of this thesis are: a speech corpus to analyze spoken dialogue in face-to-face and remote settings and an empirical analysis of conversation, collaboration, and speech prosody in both the settings. The results from the experiments show that amount of overlap is lower in remote dialogue than in the face-to-face setting. There is a significant difference in transactivity among strangers. My research benefits the computer-supported collaborative learning community by providing an analysis that can be used to build more efficient tools for supporting remote collaborative learning. / Dissertation/Thesis / Masters Thesis Computer Science 2014