Relating statistical machine learning approaches to the automatic analysis of multiparty communicative events, such as meetings, is an ambitious research area. We have investigated automatic meeting segmentation both in terms of “Meeting Actions” and “Dialogue Acts”. Dialogue acts model the discourse structure at a fine grained level highlighting individual speaker intentions. Group meeting actions describe the same process at a coarse level, highlighting interactions between different meeting participants and showing overall group intentions. A framework based on probabilistic graphical models such as dynamic Bayesian networks (DBNs) has been investigated for both tasks. Our first set of experiments is concerned with the segmentation and structuring of meetings (recorded using multiple cameras and microphones) into sequences of group meeting actions such as monologue, discussion and presentation. We outline four families of multimodal features based on speaker turns, lexical transcription, prosody, and visual motion that are extracted from the raw audio and video recordings. We relate these lowlevel multimodal features to complex group behaviours proposing a multistreammodelling framework based on dynamic Bayesian networks. Later experiments are concerned with the automatic recognition of Dialogue Acts (DAs) in multiparty conversational speech. We present a joint generative approach based on a switching DBN for DA recognition in which segmentation and classification of DAs are carried out in parallel. This approach models a set of features, related to lexical content and prosody, and incorporates a weighted interpolated factored language model. In conjunction with this joint generative model, we have also investigated the use of a discriminative approach, based on conditional random fields, to perform a reclassification of the segmented DAs. The DBN based approach yielded significant improvements when applied both to the meeting action and the dialogue act recognition task. On both tasks, the DBN framework provided an effective factorisation of the state-space and a flexible infrastructure able to integrate a heterogeneous set of resources such as continuous and discrete multimodal features, and statistical language models. Although our experiments have been principally targeted on multiparty meetings; features, models, and methodologies developed in this thesis can be employed for a wide range of applications. Moreover both group meeting actions and DAs offer valuable insights about the current conversational context providing valuable cues and features for several related research areas such as speaker addressing and focus of attention modelling, automatic speech recognition and understanding, topic and decision detection.

En cancérologie, les puces à ADN mesurant le transcriptome sont devenues un outil commun pour chercher à caractériser plus finement les pathologies, dans l’espoir de trouver au travers des expressions géniques : des mécanismes,des classes, des associations entre molécules, des réseaux d’interactions cellulaires. Ces réseaux d’interactions sont très intéressants d’un point de vue biologique car ils concentrent un grand nombre de connaissances sur le fonctionnement cellulaire. Ce travail de thèse a pour but, à partir de ces mêmes données d’expression, d’extraire des structures pouvant s’apparenter à des réseaux d’interactions génétiques. Le cadre méthodologique choisi pour appréhender cette problématique est les « Réseaux Bayésiens », c’est-à-dire une méthode à la fois graphique et probabiliste permettant de modéliser des systèmes pourtant statiques (ici le réseau d’expression génétique) à l’aide d’indépendances conditionnelles sous forme d’un réseau. L’adaptation de cette méthode à des données dont la dimension des variables (ici l’expression des gènes, dont l’ordre de grandeur est 105) est très supérieure à la dimension des échantillons (ordre102 en cancérologie) pose des problèmes statistiques (de faux positifs et négatifs) et combinatoires (avec seulement 10gènes on a 4×1018 graphes orientés sans circuit possibles). A partir de plusieurs problématiques de cancers (leucémies et cancers du sein), ce projet propose une stratégie d’accélération de recherche de réseaux d’expression à l’aide de Réseaux Bayésiens, ainsi que des mises en œuvre de cette méthode pour classer des tumeurs, sélectionner un ensemble de gènes d’intérêt reliés à une condition biologique particulière, rechercher des réseaux locaux autour d’un gène d’intérêt.On propose parallèlement de modéliser un Réseau Bayésien à partir d’un réseau biologique connu, utile pour simuler des échantillons et tester des méthodes de reconstruction de graphes à partir de données contrôlées. / In oncology, microarrays have become a classical tool to search and characterize pathologies at a deeper level than previous methods, using genetic expression to find the mechanisms, classes, molecular associations, and cellular interaction networks of different cancers. From a biological point of view, these cellular networks are interesting because they concentrate a large amount of knowledge about cellular processes. The goal of this PhD thesis project is to extract structures that could correspond to genetic interaction networks from the expression data. "Bayesian Networks", i.e. a graphic and probabilistic method that models even static systems (like the expression network) with conditional independences, are used as the framework to investigate this problem. The adaptation of this method to data where the dimension of the variables (about 105 for gene expression) is much greater than the dimension of the samples (about 102 in oncology) aggravates some statistical and combinatorial problems. For several cancer problematics, this project proposes an acceleration strategy for capturing expression networks with Bayesian Networks and some methods to classify tumors, finding gene signatures of particular biological conditions by searching for local networks in the neighborhood of a gene of interest. In parallel, we propose to model a Bayesian Network from a known biological network, which is useful to simulate samples and to test these methods to reconstruct graphs from

This dissertation defends the so-called 'agency-approach' to causation, which attempts to ground the causal relation in the cause's role of being a means to bring about its effect. The defence is confined to a conceptual interpretation of this theory, pertaining to the concept of causation as it appears in a causal judgement. However, causal judgements are not seen as limited to specific domains, and they are not exclusively attributed to human agents alone. As a methodological framework to describe the different perspectives of causal judgments, a method taken from the philosophy of information is made use of - the so-called 'method of abstraction'. According to this method, levels of abstraction are devised for the subjective perspective of the acting agent, for the agent as observer during the observation of other agents' actions, and for the agent that judges efficient causation. As a further piece of propaedeutic work, a class of similar (yet not agency-centred) approaches to causation is considered, and their modelling paradigms - Bayesian networks and interventions objectively construed - will be criticised. The dissertation then proceeds to the defence of the agency-approach, the first part of which is a defence against the objection of conceptual circularity, which holds that agency analyses causation in causal terms. While the circularity-objection is rebutted, I rely at that stage on a set of subjective concepts, i.e. concepts that are eligible to the description of the agent's own experience while performing actions. In order to give a further, positive corroboration of the agency-approach, an investigation into the natural origins and constraints of the concept of agency is made in the central chapter six of the dissertation. The thermodynamic account developed in that part affords a third-person perspective on actions, which has as its core element a cybernetic feedback cycle. At that point, the stage is set to analyse the relation between the first- and the third-person perspectives on actions previously assumed. A dual-aspect interpretation of the cybernetic-thermodynamic picture developed in chapter six will be directly applied to the levels of abstraction proposed earlier. The level of abstraction that underpins judgments of efficient causation, the kind of causation seemingly devoid of agency, will appear as a derived scheme produced by and dependent on the concept of agency. This account of efficient causation, the 'objectification of agency', affords the rebuttal of a second objection against the agency-approach, which claims that the approach is inappropriately anthropomorphic. The dissertation concludes with an account of single-case, or token level, causation, and with an examination of the impact of the causal concept on the validity of causal models.

The successful delivery of low-carbon housing (both new and retrofitted) is a key aspect of the UK s commitment to an 80% reduction in carbon emissions by 2050. In this context, the inclusion of small-scale building-integrated renewable energy technologies is an important component of low carbon design strategies, and is subject to numerous regulation and incentive schemes (including the Renewable Heat Incentive (RHI)) set up by government to encourage uptake and set minimum performance benchmarks. Unfortunately there are numerous examples of in-use energy and carbon performance shortfalls for new and retrofitted buildings this is termed the performance gap . Technical and human factors associated with building subsystem performance, which are often not considered in design tools used to predict performance, are the root cause of performance uncertainty. The research presented in this doctoral thesis aims to develop and apply a novel probabilistic method of evaluating the performance uncertainty of solar thermal systems installed in the UK. Analysis of measured data from a group of low carbon retrofitted dwellings revealed that the majority of buildings failed to meet the designed-for carbon emissions target with an average percentage difference of 60%. An in-depth case study technical evaluation of one of these dwellings showed significant dysfunction associated with the combined ASHP/solar thermal heating system, resulting in a performance gap of 94%, illustrating that the performance gap can be regarded as a whole-system problem, comprising a number of subsystem causal factors. Using a detailed dataset obtained from the UK s largest field trial of domestic solar thermal systems, a cross-cutting evaluation of predicted vs. measured performance similarly revealed a discrepancy with a mean percentage difference in predicted and measured annual yield of -24%. Having defined the nature and extent of underperformance for solar thermal technology in the UK, causal factors influencing performance were mapped and the associated uncertainty quantified using a novel knowledge-based Bayesian network (BN). In addition, the BN approach along with Monte Carlo sampling was applied to the well-established BREDEM model in order to quantify performance uncertainty of solar thermal systems by producing distributions of annual yield. As such, the modified BN-based BREDEM model represents a significant improvement in the prediction of performance of small-scale renewable energy technologies. Finally, financial analysis applied to the probabilistic predictions of annual yield revealed that the current UK RHI scheme is unlikely to result in positive returns on investment for solar thermal systems unless the duration of the payments is extended or electricity is the primary source of heating.

O segmento de fast-food tornou-se um mercado muito concorrido e com empresas bem conhecidas, tais como: Subway, McDonalds, Burguer King, Bobs e Habibs. Técnicas de inteligência artificial e ciência de dados podem oferecer inúmeros benefícios para este mercado, como por exemplo, permitir o desenvolvimento de modelos computacionais para tomada de decisões. No contexto de finanças onde envolvam a comercialização de determinados produtos, é muito comum deparar-se com cenários que envolvam incerteza, principalmente quando se deseja realizar projeções financeiras, avaliar riscos e estimativas. O objetivo deste trabalho consiste em desenvolver modelos probabilísticos baseados em Redes Bayesianas (RB) para realizar predições em vendas e análise de causalidade entre variáveis que influenciam no processo de comercialização de determinados grupos de produtos no seguimento de fast-food. Nesta análise foram avaliadas Redes Bayesianas com aprendizado de estrutura baseado em restrições, através do algoritmo Grow Shrink (GS), e Redes Bayesianas com aprendizado de estrutura baseado em pontuação, através do algoritmo Hill-Climbing (HC), posteriormente foram comparadas com um modelo de série temporal baseado em Generalized Additive Model (GAM). Os dados para análise foram adquiridos de uma rede de fast-food brasileira que possui cerca de 1100 lojas associadas, destas, foram utilizadas lojas que pertencem ao estado de São Paulo, assim como avaliado variáveis de grupos de vendas no período de 2010 à 2017. Os resultados foram avaliados através da métrica Mean Absolute Percentage Error (MAPE), que considera valores reais alimentados em modelos e valores ajustados a partir do modelo e calcula a diferença absoluta entre os dois como porcentagem do valor real, com base neste cálculo é possível obter a acurácia de cada modelo. A Rede Bayesiana (RB) com aprendizagem de estrutura baseada em pontuação, utilizando o algoritmo Hill Climbing (HC), foi escolhida como o melhor modelo, pois apresentou relações causais mais coerentes entre os vértices que influenciam o processo de venda, bem como combinações de vértices que resultam em combos de produtos, além disso, resultou em 97.60% de acurácia na previsão de vendas das lojas do estado de São Paulo (SP) na amostra de teste avaliada, com base na métrica Mean Absolute Percentage Error (MAPE). / The fast-food segment has become a busy market with well-known companies such as: Subway, McDonalds, Burger King, Bobs and Habibs. Artificial intelligence and data science techniques can offer innumerable benefits to this market, such as allowing the development of computational models for decision making. In the context of finances involving the marketing of certain products, it is very common to come across scenarios where uncertainty is involved, especially when financial projections are desired, to evaluate risks and estimation. The objective of this work is to develop probabilistic models based on Bayesian Networks (BN) to make sales predictions and causality analysis among variables that influence the commercialization process of certain product groups in the fast-food segment. In this analysis we evaluated Bayesian networks with learning of structure based on constraints, through the algorithm Grow Shrink (GS), and Bayesian Networks with learning of structure based on score, through the algorithm Hill-Climbing (HC), later were compared with a model time series based on Generalized Additive Model (GAM). The data for analysis were acquired from a Brazilian fast-food chain with approximately 1100 associated stores, of which stores were used that belong to the state of São Paulo, as well as evaluated variables of sales groups in the period from 2010 to 2017. The results were evaluated by using the Mean Absolute Percentage Error (MAPE), which considers real values fed in models and values adjusted from the model and calculates the absolute difference between the two as a percentage of the real value, based on this calculation it is possible to obtain the accuracy of each model. The Bayesian Network (BN) with scoring based structure learning, using the Hill Climbing (HC) algorithm, was chosen as the best model because it presented more coherent causal relationships between vertices that influence the sales process, as well as combinations of vertices that result in product combos, in addition, achieved a 97.60% accuracy in the sales forecast of stores in the state of Sao Paulo (SP) in the test sample evaluated, based on the Mean Absolute Percentage Error (MAPE) metric.

Redes probabilísticas são modelos muito versáteis, com aplicabilidade crescente em diversas áreas. Esses modelos são capazes de estruturar e mensurar a interação entre variáveis, permitindo que sejam realizados vários tipos de análises, desde diagnósticos de causas para algum fenômeno até previsões sobre algum evento, além de permitirem a construção de modelos de tomadas de decisões automatizadas. Neste trabalho são apresentadas as etapas para a construção dessas redes e alguns métodos usados para tal, dando maior ênfase para as chamadas redes bayesianas, uma subclasse de modelos de redes probabilísticas. A modelagem de uma rede bayesiana pode ser dividida em três etapas: seleção de variáveis, construção da estrutura da rede e estimação de probabilidades. A etapa de seleção de variáveis é usualmente feita com base nos conhecimentos subjetivos sobre o assunto estudado. A construção da estrutura pode ser realizada manualmente, levando em conta relações de causalidade entre as variáveis selecionadas, ou semi-automaticamente, através do uso de algoritmos. A última etapa, de estimação de probabilidades, pode ser feita seguindo duas abordagens principais: uma frequentista, em que os parâmetros são considerados fixos, e outra bayesiana, na qual os parâmetros são tratados como variáveis aleatórias. Além da teoria contida no trabalho, mostrando as relações entre a teoria de grafos e a construção probabilística das redes, também são apresentadas algumas aplicações desses modelos, dando destaque a problemas nas áreas de marketing e finanças. / Probabilistic networks are very versatile models, with growing applicability in many areas. These models are capable of structuring and measuring the interaction among variables, making possible various types of analyses, such as diagnoses of causes for a phenomenon and predictions about some event, besides allowing the construction of automated decision-making models. This work presents the necessary steps to construct those networks and methods used to doing so, emphasizing the so called Bayesian networks, a subclass of probabilistic networks. The Bayesian network modeling is divided in three steps: variables selection, structure learning and estimation of probabilities. The variables selection step is usually based on subjective knowledge about the studied topic. The structure learning can be performed manually, taking into account the causal relations among variables, or semi-automatically, through the use of algorithms. The last step, of probabilities estimation, can be treated following two main approaches: by the frequentist approach, where parameters are considered fixed, and by the Bayesian approach, in which parameters are treated as random variables. Besides the theory contained in this work, showing the relations between graph theory and the construction of probabilistic networks, applications of these models are presented, highlighting problems in marketing and finance.

O objeto de estudo deste trabalho é a construção de modelos que permitam a estruturação do projeto do controle de sistemas flexíveis de manufatura que considerem não somente estados de operação normal, mas também estados anormais, isto é, em situações de falhas. Entende-se como falha o desvio de pelo menos uma propriedade do sistema que leva o mesmo a um estado de defeito, que por sua vez se define como um comportamento incomum, não projetado, do sistema sob estudo, que finalmente é manifestado como um defeito. Sistemas flexíveis de manufatura são sistemas que executam múltiplos processos visando à produção de diversos bens. O processo é um conjunto de ações de transformação que por sua parte requerem um conjunto de recursos que são compartilhados por outros processos simultaneamente. Sistemas flexíveis de manufatura envolvem um número relativamente grande de componentes, máquinas, equipamentos e operadores humanos, que interagem de maneira diversificada manipulando um grande conjunto de informação e diferentes materiais em ambientes que podem até ser agressivos. Independentemente de qualquer programa de manutenção, falhas são eventos que são possíveis de acontecer em qualquer sistema de tal natureza. Num ambiente ideal, o funcionamento de todos os componentes poderia ser monitorado com o objetivo de detectar as falhas prematuras, mas devido ao custo envolvido, isso se torna inviável. Neste sentido surge o desafio de detectar as falhas a partir da observação do contexto do funcionamento do processo, mediante a monitoração de alguns parâmetros, em geral de fácil acesso, e tomando em consideração manifestações (sintomas) das falhas de um ponto de vista qualitativo. O presente trabalho propõe a utilização de redes Bayesianas para o diagnóstico de falhas em sistemas flexíveis de manufatura. As redes Bayesianas constituem uma ferramenta útil para a representação das relações que existem entre as causas (componentes em estado de falha) e os sintomas (observações anormais do processo). A partir deste modelo, inferências podem ser feitas para o diagnóstico do sistema.Por outro lado, nos últimos anos a rede de Petri tem sido utilizada exitosamente na representação dos aspectos de controle de sistemas produtivos e particularmente de sistemas de manufatura e, desta forma, considera-se aqui tal ferramenta para a modelagem do sistema não só em condições normais de funcionamento como também para a representação do tratamento de falhas, no contexto de um sistema tolerante a anomalias do processo. Especial ênfase é dada à estruturação de uma metodologia que permita a concepção de um procedimento eficaz para a construção de modelos de controle. / The objective of the present work is the construction of models proper for the easy implementation of flexible manufacturing control systems able to handle not only with normal behavioral conditions, but with abnormal (or faulty) behavior as well. A fault is defined as a deviation of at least one system property that drives the system into an error state. An error is defined as an uncommon behavior, not expected from the system functionalities. Flexible manufacturing systems are systems that execute multiple processes for the production of several items in several ways. A process is a sequence of certain transformation tasks that require a set of resources shared simultaneously by multiple processes. In this sense, flexible manufacturing systems are constituted of a relatively great number of devices, machines, equipments and human operators that work together manipulating great quantities of information and materials. This work is usually performed in aggressive environments. So, independent of any maintenance program, faults are events that cannot be totally avoided. In an ideal environment, the monitoring of all components is the way to avoid faults. Nevertheless, due to the cost involved, this is an impossible task. In this context, there is a challenge to properly detect faults from the observation of the systems context, through the monitoring and observation of some parameters in general easy to access, including also qualitative information from operators. In the present work, it is proposed the use of Bayesian networks for the fault diagnosis in flexible manufacturing systems. Bayesian networks constitute a useful tool for the modeling of the causal relation between the causes (faulty components) and the symptoms (manifestations). Based on this model, inference can be done for the system diagnosis task. Additionally, in the last years Petri net has been successfully used for the modeling of control systems of productive systems and particularly, manufacturing control systems. In this work, beyond the use of Petri net for the modeling of normal situations of the system, Petri net is used for the modeling of the fault treatment techniques. This drives the system tolerance to faults. Especial emphasis is laid into methodological issues that allows for the structuration of a systematic procedure proper for the modeling and construction of control systems.

O presente trabalho tem como objetivo principal a classificação do risco de infestação por regiões de culturas vegetais por plantas daninhas. Os riscos por regiões são obtidos por um sistema de classificação fuzzy, usando métodos de Krigagem e análise de imagens. A infestação é descrita por atributos da cobertura foliar, densidade de sementes, extensão dos agrupamentos de sementes e competitividade, obtidos a partir das amostras de densidades de sementes e de plantas daninhas, da cobertura foliar e da biomassa de plantas daninhas. O atributo da cobertura foliar indica a porcentagem de ocupação das plantas emergentes e é obtido a partir de um mapa de cobertura foliar, construído usando Krigagem. O atributo da densidade de sementes caracteriza a localização das sementes que podem germinar e é obtido a partir de um mapa da distribuição da produção de sementes das plantas daninhas, também construído usando Krigagem. O atributo da extensão dos agrupamentos de sementes reflete a influência das sementes vizinhas em uma certa localização e também é obtido a partir do mapa de distribuição da produção de sementes. O atributo da competitividade entre plantas daninhas e cultura é obtido a partir de um sistema neurofuzzy, utilizando amostras de densidade e de biomassa das plantas daninhas. Para reunir os riscos de infestação semelhantes, os valores de risco inferidos por região pelo sistema fuzzy são agrupados considerando valores e localizações próximas utilizando o método k-médias com coeficiente de variação. Uma abordagem probabilística com redes de classificação Bayesianas é também empregada para a obtenção de um conjunto de regras linguísticas para classificar a competitividade e o risco de infestação, por motivo de comparação. Resultados para o risco de infestação são obtidos para uma área experimental em uma cultura de milho indicando a existência de riscos diferenciados que são explicados pela perda de rendimento da cultura. / The goal of this work is the classification of the risk of infestation per regions of a crop by weeds. The risks per regions are obtained by a fuzzy classification system, using kriging and image analysis. The infestation is described by attributes of the weed coverage, weed seed density, weed seed patches and competitiveness, obtained from weed seeds and weed densities, weed coverage and biomass. The attribute of the weed coverage indicates the percentage of infested surface of the emergent weeds which is obtained from a weed coverage map built with kriging. The attribute of the weed seed density is obtained from a weed seed production map also built with kriging which characterizes the locations of seeds which can germinate. The attribute of the weed seed patches is also obtained by the weed seed production map which reflects how the seeds contribute to weed proliferation in the surroundings. The attribute of the competitiveness among weeds and crop is obtained from a neurofuzzy system, using the weeds density and biomass of the plants. In order to aggregate the similar risks of infestation, the values of risks per region inferred by the fuzzy system are clustered according to similar values and locations using the k-means method with a variation coefficient. A probabilistic approach with Bayesian networks classifiers is also considered to obtain a set of linguistic rules to classify the competitiveness and the risk of infestation, for comparison purposes. Results for the risk of infestation are obtained for an experimental area in a corn crop which indicate the existence of different risks, explained by the yield loss of the crop.

Les travaux menés dans cette thèse CIFRE s’inscrivent dans le cadre d’une collaboration pérenne entre le CRAN et l'EDF R&D dont un des résultats majeurs a été le développement d'une méthodologie d’analyse de risques, appelée Analyse Intégrée des Risques (AiDR). Cette méthodologie traite des systèmes sociotechniques sous les angles technique, humain et organisationnel et dont les équipements sont soumis à des actions de maintenance et/ou de conduite. La thèse a pour objet ainsi de proposer une évolution du modèle dit de « barrière humaine » développé dans l'AiDR pour évaluer l'efficacité de ces actions humaines prises leur contexte organisationnel. Nos contributions majeures s'organisent autour de 3 axes : 1. Une amélioration de la structure préexistante du modèle de barrière humaine afin d’aboutir à un modèle basé sur des facteurs de forme appelés performance shaping factors (PSF) fournis par les méthodes d’Évaluation Probabiliste de la Fiabilité Humaine (EPFH) ;2. L’intégration de la résilience et la modélisation de l’interaction entre mécanismes résilients et pathogènes impactant l'efficacité des actions dans les relations causales probabilistes ;3. Un traitement global des jugements d’expert cohérent avec la structure mathématique du modèle proposé permettant d’estimer d’une manière objective les paramètres du modèle. Ce traitement se fonde sur la construction d’un questionnaire permettant de "guider" l’expert vers l’évaluation d’effets conjoints issus de l’interaction entre mécanismes pathogènes et résilients. L’ensemble des contributions proposées a été validé sur un cas d’application portant sur une barrière humaine mise en place dans un cas d’inondation externe d’une unité de production d’électricité d’EDF / The work carried out in this CIFRE PhD thesis is part of a long-term collaboration between CRAN and EDF R&D, one of the major results of which was the development of a risk analysis methodology called Integrated Risk Analysis (AiDR). This methodology deals with sociotechnical systems from technical, human and organizational points of view and whose equipment is subjected to maintenance and/or operation activities. This thesis aims to propose an evolution of the so-called "human barrier" model developed in the AiDR in order to evaluate the effectiveness of these human actions taken their organizational context. Our major contributions are organized around 3 axes: 1. Improvement of the pre-existing structure of the human barrier model to achieve a model based on performance shaping factors (PSF) provided by the Human Reliability Assessment (HRA) methods; 2. Integration of resilience and modeling of the interaction between resilient and pathogenic mechanisms impacting the effectiveness of activities in a probabilistic causal framework; 3. A global treatment of the expert judgments consistent with the mathematical structure of the proposed model in order to objectively estimate the parameters of the model. This treatment is based on a questionnaire to guide experts towards the evaluation of joint effects resulting from the interaction between pathogenic and resilient mechanisms. All of the proposed contributions have been validated on an application case involving a human barrier put in place during an external flooding occurring at an EDF power plant

Made available in DSpace on 2017-07-21T14:19:22Z (GMT). No. of bitstreams: 1 Valter Luis.pdf: 3516533 bytes, checksum: d498d5c67dd1b9a837a128c20cabef67 (MD5) Previous issue date: 2015-02-26 / The interestingness area of data mining process aiming to reduce the amount of models to be analyzed for experts in the interpretation step of the knowledge discovery in databases. In this work, a method for analysis the interestingness of regression models was developed. This method combine probabilistic multivariate models with Pearson correlation test and Wilcoxon signed-rank test resulting in a new interestingness measure, named Impact. The developed method was applied over regression models found during a data mining process for estimating agricultural gypsum requirements. The results showed that the probabilistic multivariate filter was able to filter the best models according to a utility-based approach, in this case, for practical application on agriculture. Six models were considered interesting, with Impact score > 0.5, and only one was miscategorized. On the other hand, the combined statistical test filters were able to filter six models two of them were miscategorized. The attributes identified as most relevant to estimate gypsum rate were: time, Ca and its concentration on effective cation exchange capacity (CaCTCe), mainly in superficial layers. / A interessabilidade de regras é uma área da mineração de dados que tem por objetivo reduzir a quantidade de modelos a serem analisados por especialistas na etapa de interpretação do conhecimento descoberto em bases de dados. Embora existam várias medidas de interesse de regras voltadas para as tarefas de associação e classificação, observa-se uma falta de métodos consolidados para análise de interessabilidade de modelos de regressão. Neste trabalho foi desenvolvido um método para analisar a interessabilidade de modelos de regressão, o qual combina um filtro baseado em modelos probabilísticos multivariados com filtros baseados em testes estatísticos de correlação de Pearson e de postos de sinais de Wilcoxon, resultando em uma nova medida de interessabilidade denominada Impacto. O método desenvolvido foi aplicado sobre modelos de regressão encontrados no processo de mineração de dados para estimativa de gesso agrícola. Estes dados resultam de três experimentos sob Sistema Plantio Direto realizados na Região dos Campos Gerais, PR, nos quais foram medidos, em diferentes épocas, os teores dos nutrientes do solo após a aplicação de doses de gesso. Os resultados mostraram que o filtro probabilístico multivariado foi capaz de filtrar os melhores modelos segundo uma visão de utilidade, ou seja, de potencial de aplicação agronômica. Foram selecionados seis modelos com score de Impacto > 0,5, ou seja, considerados interessantes, e destes apenas um foi considerado incorretamente classificado. Por outro lado, os filtros baseados em testes estatísticos foram capazes de filtrar seis modelos sendo que dois deles podem ser considerados incorretamente classificados. Os atributos identificados como mais relevantes para o problema do gesso agrícola foram a época, o teor de Ca e a concentração de Ca em relação à capacidade de troca catiônica efetiva (CTCe), especialmente em camadas superficiais do solo.