This paper uses statistical learning to examine and compare three different statistical methods with the aim to predict credit card fraud. The methods compared are Logistic Regression, K-Nearest Neighbour and Random Forest. They are applied and estimated on a data set consisting of nearly 300,000 credit card transactions to determine their performance using classification of fraud as the outcome variable. The three models all have different properties and advantages. The K-NN model preformed the best in this paper but has some disadvantages, since it does not explain the data but rather predict the outcome accurately. Random Forest explains the variables but performs less precise. The Logistic Regression model seems to be unfit for this specific data set.

O principal objetivo do estudo foi utilizar informações de ocorrência do Traumatismo Crânio Encefálico (TCE) que possam inferir/gerar descobertas associadas ao risco de gravidade do paciente, bem como auxiliar na tomada de decisão médica ao definir o melhor prognóstico, indicando quais as possíveis medidas que podem ser escolhidas para a gravidade na lesão sofrida pela vítima. Inicialmente, foram analisadas as estatísticas descritivas dos dados dos pacientes de TCE de um hospital do interior de São Paulo. Participaram desse estudo 50 pacientes. Os resultados mostraram que a maior frequência do trauma é por acidentes de trânsito (62%), seguidos de acidentes por queda (24%). Traumas em pacientes do sexo masculino (88%) são muito mais frequentes do que em pacientes do sexo feminino. Para modelagem, transformou-se a variável resposta \"Abbreviated Injury Scale (AIS)\" em dicotômica, considerando 0 (zero) aos pacientes fora de risco e 1 (um) aos que apresentaram algum tipo de risco. Em seguida, técnicas de aprendizado estatístico foram utilizadas de modo a comparar o desempenho dos classificadores Regressão Logística sendo um caso do Generalized Linear Model (GLM), Random Forest (RF), Support Vector Machine (SVM) e redes probabilísticas Naïve Bayes (NB). O modelo com melhor desempenho (RF) combinou os índices Accuracy (ACC) , Area Under ROC Curve (AUC) , Sensitivity (SEN), Specificity (SPE) e Matthews Correlation Coefficient (MCC), que apresentaram os resultados mais favoráveis no quesito de apoio no auxílio da tomada de decisão médica, possibilitando escolher o estudo clínico mais adequado das vítimas traumatizadas ao considerar o risco de vida do indivíduo. Conforme o modelo selecionado foi possível gerar um ranking para estimar a probabilidade de risco de vida do paciente. Em seguida foi realizado uma comparação de desempenho entre o modelo RF (novo classificador) e os índices Revisited Trauma Score (RTS), Injury Severity Score (ISS) , Índice de Barthel (IB) referente à classificação de risco dos pacientes. / The main objective of this study was to consider the information related to the occurrence of traumatic brain injury (TBI) that can infer new results associated with the patients risk of severity as well as assisting in the medical decision in order to find the best prognosis; this can lead to indicate possible measures that can be chosen for severity in the injury suffered by the victim. Initially, we have presented descriptive statistics from the patients with TBI from a hospital located in the heartland of São Paulo. Fifty patients were recruited for this study. Descriptive analyzes showed that the highest frequency of trauma is due to traffic accidents (62 %) followed by crashes per accident (24 %). The causes related to trauma occur much more often in male patients (88 %) than in female patients. To order model, the response variable Abbreviated Injury Scale (AIS) was considered as dichotomous, where 0 (zero) was to out-of-risk patients and 1 (one) to those who presented some type of risk. Further, statistical learning techniques were used in order to compare the performance of the Logistic Regression as a Generalized Linear Model (GLM), Random Forest (RF), Support Vector Machine (SVM) and Naive Bayes (NB) model. The best performing (RF) model combined the Accuracy (ACC) , Area Under ROC Curve (AUC) , Sensitivity (SEN), Specificity (SPE) e Matthews Correlation Coefficient (MCC), which presented the most favorable results in terms of support in medical decision, making it possible to choose the most appropriate clinical study of traumatized victims based on the individual life risk. According to the selected model it was possible to generate a rank to estimate the probability of life risk of the patient. Then a performance comparison was performed between the RF model (proposed classifier) and the Revisited Trauma Score (RTS), Injury Severity Score (ISS), Barthel index (IB) referring to the risk classification of patients.

De nombreux algorithmes en Apprentissage Automatique utilisent une notion de distance ou de similarité entre les exemples pour résoudre divers problèmes tels que la classification, le partitionnement ou l'adaptation de domaine. En fonction des tâches considérées ces métriques devraient avoir des propriétés différentes mais les choisir manuellement peut-être fastidieux et difficile. Une solution naturelle est alors d'adapter automatiquement ces métriques à la tâche considérée. Il s'agit alors d'un problème connu sous le nom d'Apprentissage de Métriques et où le but est principalement de trouver les meilleurs paramètres d'une métrique respectant des contraintes spécifiques. Les approches classiques dans ce domaine se focalisent habituellement sur l'apprentissage de distances de Mahalanobis ou de similarités bilinéaires et l'une des principales limitations est le fait que le contrôle du comportement de ces métriques est souvent limité. De plus, si des travaux théoriques existent pour justifier de la capacité de généralisation des modèles appris, la plupart des approches ne présentent pas de telles garanties. Dans cette thèse nous proposons de nouveaux algorithmes pour apprendre des métriques à comportement contrôlé et nous mettons l'accent sur les propriétés théoriques de ceux-ci. Nous proposons quatre contributions distinctes qui peuvent être séparées en deux parties: (i) contrôler la métrique apprise en utilisant une métrique de référence et (ii) contrôler la transformation induite par la métrique apprise. Notre première contribution est une approche locale d'apprentissage de métriques où le but est de régresser une distance proportionnelle à la perception humaine des couleurs. Notre approche est justifiée théoriquement par des garanties en généralisation sur les métriques apprises. Dans notre deuxième contribution nous nous sommes intéressés à l'analyse théorique de l'intérêt d'utiliser une métrique de référence dans un terme de régularisation biaisé pour aider lors du processus d'apprentissage. Nous proposons d'utiliser trois cadres théoriques différents qui nous permettent de dériver trois mesures différentes de l'apport de la métrique de référence. Ces mesures nous donnent un aperçu de l'impact de la métrique de référence sur celle apprise. Dans notre troisième contribution nous proposons un algorithme d'apprentissage de métriques où la transformation induite est contrôlée. L'idée est que, plutôt que d'utiliser des contraintes de similarité et de dissimilarité, chaque exemple est associé à un point virtuel qui appartient déjà à l'espace induit par la métrique apprise. D'un point de vue théorique nous montrons que les métriques apprises de cette façon généralisent bien mais aussi que notre approche est liée à une méthode plus classique d'apprentissage de métriques basée sur des contraintes de paires. Dans notre quatrième contribution nous essayons aussi de contrôler la transformation induite par une métrique apprise. Cependant, plutôt que considérer un contrôle individuel pour chaque exemple, nous proposons une approche plus globale en forçant la transformation à suivre une transformation géométrique associée à un problème de transport optimal. D'un point de vue théorique nous proposons une discussion sur le lien entre la transformation associée à la métrique apprise et la transformation associée au problème de transport optimal. D'un point de vue plus pratique nous montrons l'intérêt de notre approche pour l'adaptation de domaine mais aussi pour l'édition d'images / Many Machine Learning algorithms make use of a notion of distance or similarity between examples to solve various problems such as classification, clustering or domain adaptation. Depending on the tasks considered these metrics should have different properties but manually choosing an adapted comparison function can be tedious and difficult. A natural trend is then to automatically tailor such metrics to the task at hand. This is known as Metric Learning and the goal is mainly to find the best parameters of a metric under some specific constraints. Standard approaches in this field usually focus on learning Mahalanobis distances or Bilinear similarities and one of the main limitations is that the control over the behaviour of the learned metrics is often limited. Furthermore if some theoretical works exist to justify the generalization ability of the learned models, most of the approaches do not come with such guarantees. In this thesis we propose new algorithms to learn metrics with a controlled behaviour and we put a particular emphasis on the theoretical properties of these algorithms. We propose four distinct contributions which can be separated in two parts, namely (i) controlling the metric with respect to a reference metric and (ii) controlling the underlying transformation corresponding to the learned metric. Our first contribution is a local metric learning method where the goal is to regress a distance proportional to the human perception of colors. Our approach is backed up by theoretical guarantees on the generalization ability of the learned metrics. In our second contribution we are interested in theoretically studying the interest of using a reference metric in a biased regularization term to help during the learning process. We propose to use three different theoretical frameworks allowing us to derive three different measures of goodness for the reference metric. These measures give us some insights on the impact of the reference metric on the learned one. In our third contribution we propose a metric learning algorithm where the underlying transformation is controlled. The idea is that instead of using similarity and dissimilarity constraints we associate each learning example to a so-called virtual point belonging to the output space associated with the learned metric. We theoretically show that metrics learned in this way generalize well but also that our approach is linked to a classic metric learning method based on pairs constraints. In our fourth contribution we also try to control the underlying transformation of a learned metric. However instead of considering a point-wise control we consider a global one by forcing the transformation to follow the geometrical transformation associated to an optimal transport problem. From a theoretical standpoint we propose a discussion on the link between the transformation associated with the learned metric and the transformation associated with the optimal transport problem. On a more practical side we show the interest of our approach for domain adaptation but also for a task of seamless copy in images

A novelty detection algorithm inspired by human audio pattern recognition is conceptualized and experimentally tested. This anomaly detection technique can be used to monitor the health of a machine or could also be coupled with a current state of the art system to enhance its fault detection capabilities. Time-domain data obtained from a microphone is processed by applying a short-time FFT, which returns time-frequency patterns. Such patterns are fed to a machine learning algorithm, which is designed to detect novel signals and identify windows in the frequency domain where such novelties occur. The algorithm presented in this paper uses one-dimensional kernel density estimation for different frequency bins. This process eliminates the need for data dimension reduction algorithms. The method of "pseudo-likelihood cross validation" is used to find an independent optimal kernel bandwidth for each frequency bin. Metrics such as the "Individual Node Relative Difference" and "Total Novelty Score" are presented in this work, and used to assess the degree of novelty of a new signal. Experimental datasets containing synthetic and real novelties are used to illustrate and test the novelty detection algorithm. Novelties are successfully detected in all experiments. The presented novelty detection technique could greatly enhance the performance of current state-of-the art condition monitoring systems, or could also be used as a stand-alone system.

In a world confronting climate change and rapidly shifting land uses, effective methods for monitoring natural resources are critical to support scientifically-informed management decisions. By taking audio recordings of the environment, scientists can acquire presence-absence data to characterize populations of sound-producing wildlife over time and across vast spatial scales. Remote acoustic monitoring presents new challenges, however: monitoring programs are often constrained in the total time they can record, automated detection algorithms typically produce a prohibitive number of detection mistakes, and there is no streamlined framework for moving from raw acoustic data to models of wildlife occurrence dynamics. In partnership with a proof-of-concept field study in the U.S Bureau of Land Management’s Riverside East Solar Energy Zone in southern California, this dissertation introduces a new R software package, AMMonitor, alongside a novel body of work: 1) temporally-adaptive acoustic sampling to maximize the detection probabilities of target species despite recording constraints, 2) values-driven statistical learning tools for template-based automated detection of target species, and 3) methods supporting the construction of dynamic species occurrence models from automated acoustic detection data. Unifying these methods with streamlined data management, the AMMonitor software package supports the tracking of species occurrence, colonization, and extinction patterns through time, introducing the potential to perform adaptive management at landscape scales.

With emergence of biomedical informatics, Web intelligence, and E-business, new challenges are coming for knowledge discovery and data mining modeling problems. In this dissertation work, a framework named Granular Support Vector Machines (GSVM) is proposed to systematically and formally combine statistical learning theory, granular computing theory and soft computing theory to address challenging predictive data modeling problems effectively and/or efficiently, with specific focus on binary classification problems. In general, GSVM works in 3 steps. Step 1 is granulation to build a sequence of information granules from the original dataset or from the original feature space. Step 2 is modeling Support Vector Machines (SVM) in some of these information granules when necessary. Finally, step 3 is aggregation to consolidate information in these granules at suitable abstract level. A good granulation method to find suitable granules is crucial for modeling a good GSVM. Under this framework, many different granulation algorithms including the GSVM-CMW (cumulative margin width) algorithm, the GSVM-AR (association rule mining) algorithm, a family of GSVM-RFE (recursive feature elimination) algorithms, the GSVM-DC (data cleaning) algorithm and the GSVM-RU (repetitive undersampling) algorithm are designed for binary classification problems with different characteristics. The empirical studies in biomedical domain and many other application domains demonstrate that the framework is promising. As a preliminary step, this dissertation work will be extended in the future to build a Granular Computing based Predictive Data Modeling framework (GrC-PDM) with which we can create hybrid adaptive intelligent data mining systems for high quality prediction.

The study of network-service disruptions caused by large-scale disturbances has mainly focused on assessing network damage; however, network-disruption responses, i.e., how the disruptions occur depending on social organizations, weather, and power resources, have been studied little. The goal of this research is to study the responses of network-service disruptions caused by large-scale disturbances with respect to (1) temporal and logical network, and (2) external factors such as weather and power resources, using real and publicly available heterogeneous data that are composed of network measurements, user inputs, organizations, geographic locations, weather, and power outage reports. Network-service disruptions at the subnet level caused by Hurricanes Katrina in 2005 and Ike in 2008 are used as the case studies. The analysis of network-disruption responses with respect to temporal and logical network shows that subnets became unreachable dependently within organization, cross organization, and cross autonomous system. Thus, temporal dependence also illustrates the characteristics of logical dependence. In addition, subnet unreachability is analyzed with respect to the external factors. It is found that subnet unreachability and the storm are weakly correlated. The weak correlation motivates us to search for root causes and discover that the majority of subnet unreachability reportedly occurred because of power outages or lack of power generators. Using the power outage data, it is found that subnet unreachability and power outages are strongly correlated.

Remote sensing techniques are vital for early detection of several problems such as natural disasters, ecological problems and collecting information necessary for finding optimum solutions to those problems. Remotely sensed information has also important uses in predicting the future risks, urban planning, communication.Recent developments in remote sensing instrumentation offered a challenge to the mathematical and statistical methods to process the acquired information. Classification of satellite images in the context of land cover classification is the main concern of this study. Land cover classification can be performed by statistical learning methods like additive models, decision trees, neural networks, k-means methods which are already popular in unsupervised classification and clustering of image scene inverse problems. Due to the degradation and corruption of satellite images, the classification performance is limited both by the accuracy of clustering and by the extent of the classification. In this study, we are concerned with understanding the performance of the available unsupervised methods with k-means, supervised methods with Gaussian maximum likelihood which are very popular methods in land cover classification. A broader approach to the classification problem based on finding the optimal discriminants from a larger range of functions is considered also in this work. A novel method based on threshold decomposition and Boolean discriminant functions is developed as an implementable application of this approach. All methods are applied to BILSAT and Landsat satellite images using MATLAB software.

The aim of this paper is to investigate if rainfall prediction (nowcasting) can successively be made using a deep learning approach. The input to the networks are different spatiotemporal variables including forecasts from a NWP model. The results indicate that these networks has some predictive power and could be use in real application. Another interesting empirical finding relates to the usage of transfer learning from a domain which is not related instead of random initialization. Using pretrained parameters resulted in better convergence and overall performance than random initialization of the parameters.

La notion de métrique joue un rôle clef dans les problèmes d’apprentissage automatique tels que la classification, le clustering et le ranking. L’apprentissage à partir de données de métriques adaptées à une tâche spécifique a suscité un intérêt croissant ces dernières années. Ce domaine vise généralement à trouver les meilleurs paramètres pour une métrique donnée sous certaines contraintes imposées par les données. La métrique apprise est utilisée dans un algorithme d’apprentissage automatique dans le but d’améliorer sa performance. La plupart des méthodes d’apprentissage de métriques optimisent les paramètres d’une distance de Mahalanobis pour des vecteurs de features. Les méthodes actuelles de l’état de l’art arrivent à traiter des jeux de données de tailles significatives. En revanche, le sujet plus complexe des séries temporelles multivariées n’a reçu qu’une attention limitée, malgré l’omniprésence de ce type de données dans les applications réelles. Une importante partie de la recherche sur les séries temporelles est basée sur la dynamic time warping (DTW), qui détermine l’alignement optimal entre deux séries temporelles. L’état actuel de l’apprentissage de métriques souffre de certaines limitations. La plus importante est probablement le manque de garanties théoriques concernant la métrique apprise et sa performance pour la classification. La théorie des fonctions de similarité (ℰ , ϓ, T)-bonnes a été l’un des premiers résultats liant les propriétés d’une similarité à celles du classifieur qui l’utilise. Une deuxième limitation vient du fait que la plupart des méthodes imposent des propriétés de distance, qui sont coûteuses en terme de calcul et souvent non justifiées. Dans cette thèse, nous abordons les limitations précédentes à travers deux contributions principales. La première est un nouveau cadre général pour l’apprentissage conjoint d’une fonction de similarité et d’un classifieur linéaire. Cette formulation est inspirée de la théorie de similarités (ℰ , ϓ, τ) -bonnes, fournissant un lien entre la similarité et le classifieur linéaire. Elle est convexe pour une large gamme de fonctions de similarité et de régulariseurs. Nous dérivons deux bornes de généralisation équivalentes à travers les cadres de robustesse algorithmique et de convergence uniforme basée sur la complexité de Rademacher, prouvant les propriétés théoriques de notre formulation. Notre deuxième contribution est une méthode d’apprentissage de similarités basée sur DTW pour la classification de séries temporelles multivariées. Le problème est convexe et utilise la théorie des fonctions (ℰ , ϓ, T)-bonnes liant la performance de la métrique à celle du classifieur linéaire associé. A l’aide de la stabilité uniforme, nous prouvons la consistance de la similarité apprise conduisant à la dérivation d’une borne de généralisation. / The notion of metric plays a key role in machine learning problems, such as classification, clustering and ranking. Learning metrics from training data in order to make them adapted to the task at hand has attracted a growing interest in the past years. This research field, known as metric learning, usually aims at finding the best parameters for a given metric under some constraints from the data. The learned metric is used in a machine learning algorithm in hopes of improving performance. Most of the metric learning algorithms focus on learning the parameters of Mahalanobis distances for feature vectors. Current state of the art methods scale well for datasets of significant size. On the other hand, the more complex topic of multivariate time series has received only limited attention, despite the omnipresence of this type of data in applications. An important part of the research on time series is based on the dynamic time warping (DTW) computing the optimal alignment between two time series. The current state of metric learning suffers from some significant limitations which we aim to address in this thesis. The most important one is probably the lack of theoretical guarantees for the learned metric and its performance for classification.The theory of (ℰ , ϓ, τ)-good similarity functions has been one of the first results relating the properties of a similarity to its classification performance. A second limitation in metric learning comes from the fact that most methods work with metrics that enforce distance properties, which are computationally expensive and often not justified. In this thesis, we address these limitations through two main contributions. The first one is a novel general framework for jointly learning a similarity function and a linear classifier. This formulation is inspired from the (ℰ , ϓ, τ)-good theory, providing a link between the similarity and the linear classifier. It is also convex for a broad range of similarity functions and regularizers. We derive two equivalent generalization bounds through the frameworks of algorithmic robustness and uniform convergence using the Rademacher complexity, proving the good theoretical properties of our framework. Our second contribution is a method for learning similarity functions based on DTW for multivariate time series classification. The formulation is convex and makes use of the(ℰ , ϓ, τ)-good framework for relating the performance of the metric to that of its associated linear classifier. Using uniform stability arguments, we prove the consistency of the learned similarity leading to the derivation of a generalization bound.