Frequency Hopping Spread Spectrum (FHSS) signaling is used across many devices operating in both regulated and unregulated bands. In either situation, if there is a malicious device operating within these bands, or more simply a user operating out of the required specifications, the identification this user important to insure communication link integrity and interference mitigation. The identification of a user involves the grouping of that users signal transmissions, and the separation of those transmission from transmissions of other users in a shared frequency band. Traditional signal separation methods often require difficult to obtain hardware fingerprinting characteristics or approximate geo-location estimates. This work will consider the characteristics of FHSS signals that can be extracted directly from signal detection. From estimates of these hopping characteristics, novel source separation with classic clustering algorithms can be performed. Background knowledge derived from the time domain representation of received waveforms can improve these clustering methods with the novel application of cannot-link pairwise constraints to signal separation. For equivalent clustering accuracy, constraint-based clustering tolerates higher parameter estimation error, caused by diminishing received signal-to-noise ratio (SNR), for example. Additionally, prior work does not fully cover the implications of detecting and estimating FHSS signals using image segmentation on a Time-Frequency (TF) waterfall. This work will compare several methods of FHSS signal detection, and discuss how each method may effect estimation accuracy and signal separation quality. The use of constraint-based clustering is shown to provide higher clustering accuracy, resulting in more reliable separation and identification of active users in comparison to traditional clustering methods. / Master of Science / The expansion of technology in areas such as smart homes and appliances, personal devices, smart vehicles, and many others, leads to more and more devices using common wireless communication techniques such as WiFi and Bluetooth. While the number of wirelessly connected users expands, the range of frequencies that support wireless communications does not. It is therefore essential that each of these devices unselfishly share the available wireless resources. If a device is using more resources than the required limits, or causing interference with other’s communications, this device will impact many others negatively and therefore preventative action must be taken to prevent further disruption in the wireless environment. Before action can be taken however, the device must first be identified in a mixture of other wireless activity. To identify a specific device, first, a wireless receiver must be in close enough proximity to detect the power that the malicious device is emitting through its wireless communication. This thesis provides a method that can be used to identify a problem user based only off of its wireless transmission behavior. The performance of this identification is shown with respect to the received signal power which represents the necessary range that a listening device must be to identify and separate a problem user from other cooperative users that are communicating wirelessly.

La classification non supervisée, souvent appelée par le terme anglais de clustering, est une tâche importante en Fouille de Données. Depuis une dizaine d'années, la classification non supervisée a été étendue pour intégrer des contraintes utilisateur permettant de modéliser des connaissances préalables dans le processus de clustering. Différents types de contraintes utilisateur peuvent être considérés, des contraintes pouvant porter soit sur les clusters, soit sur les instances. Dans cette thèse, nous étudions le cadre de la Programmation par Contraintes (PPC) pour modéliser les tâches de clustering sous contraintes utilisateur. Utiliser la PPC a deux avantages principaux : la déclarativité, qui permet d'intégrer aisément des contraintes utilisateur et la capacité de trouver une solution optimale qui satisfait toutes les contraintes (s'il en existe). Nous proposons deux modèles basés sur la PPC pour le clustering sous contraintes utilisateur. Les modèles sont généraux et flexibles, ils permettent d'intégrer des contraintes d'instances must-link et cannot-link et différents types de contraintes sur les clusters. Ils offrent également à l'utilisateur le choix entre différents critères d'optimisation. Afin d'améliorer l'efficacité, divers aspects sont étudiés. Les expérimentations sur des bases de données classiques et variées montrent qu'ils sont compétitifs par rapport aux approches exactes existantes. Nous montrons que nos modèles peuvent être intégrés dans une procédure plus générale et nous l'illustrons par la recherche de la frontière de Pareto dans un problème de clustering bi-critère sous contraintes utilisateur. / Cluster analysis is an important task in Data Mining with hundreds of different approaches in the literature. Since the last decade, the cluster analysis has been extended to constrained clustering, also called semi-supervised clustering, so as to integrate previous knowledge on data to clustering algorithms. In this dissertation, we explore Constraint Programming (CP) for solving the task of constrained clustering. The main principles in CP are: (1) users specify declaratively the problem in a Constraint Satisfaction Problem; (2) solvers search for solutions by constraint propagation and search. Relying on CP has two main advantages: the declarativity, which enables to easily add new constraints and the ability to find an optimal solution satisfying all the constraints (when there exists one). We propose two models based on CP to address constrained clustering tasks. The models are flexible and general and supports instance-level constraints and different cluster-level constraints. It also allows the users to choose among different optimization criteria. In order to improve the efficiency, different aspects have been studied in the dissertation. Experiments on various classical datasets show that our models are competitive with other exact approaches. We show that our models can easily be embedded in a more general process and we illustrate this on the problem of finding the Pareto front of a bi-criterion optimization process.

The main focus of my research is to design effective learning techniques for information retrieval and mining in high-dimensional databases. There are two main aspects in the retrieval and mining research: accuracy and efficiency. The accuracy problem is how to return results which can better match the ground truth, and the efficiency problem is how to evaluate users' requests and execute learning algorithms as fast as possible. However, these problems are non-trivial because of the complexity of the high-level semantic concepts, the heterogeneous natures of the feature space, the high dimensionality of data representations and the size of the databases. My dissertation is dedicated to addressing these issues. Specifically, my work has five main contributions as follows. The first contribution is a novel manifold learning algorithm, Local and Global Structures Preserving Projection (LGSPP), which defines salient low-dimensional representations for the high-dimensional data. A small number of projection directions are sought in order to properly preserve the local and global structures for the original data. Specifically, two groups of points are extracted for each individual point in the dataset: the first group contains the nearest neighbors of the point, and the other set are a few sampled points far away from the point. These two point sets respectively characterize the local and global structures with regard to the data point. The objective of the embedding is to minimize the distances of the points in each local neighborhood and also to disperse the points far away from their respective remote points in the original space. In this way, the relationships between the data in the original space are well preserved with little distortions. The second contribution is a new constrained clustering algorithm. Conventionally, clustering is an unsupervised learning problem, which systematically partitions a dataset into a small set of clusters such that data in each cluster appear similar to each other compared with those in other clusters. In the proposal, the partial human knowledge is exploited to find better clustering results. Two kinds of constraints are integrated into the clustering algorithm. One is the must-link constraint, indicating that the involved two points belong to the same cluster. On the other hand, the cannot-link constraint denotes that two points are not within the same cluster. Given the input constraints, data points are arranged into small groups and a graph is constructed to preserve the semantic relations between these groups. The assignment procedure makes a best effort to assign each group to a feasible cluster without violating the constraints. The theoretical analysis reveals that the probability of data points being assigned to the true clusters is much higher by the new proposal, compared to conventional methods. In general, the new scheme can produce clusters which can better match the ground truth and respect the semantic relations between points inferred from the constraints. The third contribution is a unified framework for partition-based dimension reduction techniques, which allows efficient similarity retrieval in the high-dimensional data space. Recent similarity search techniques, such as Piecewise Aggregate Approximation (PAA), Segmented Means (SMEAN) and Mean-Standard deviation (MS), prove to be very effective in reducing data dimensionality by partitioning dimensions into subsets and extracting aggregate values from each dimension subset. These partition-based techniques have many advantages including very efficient multi-phased pruning while being simple to implement. They, however, are not adaptive to different characteristics of data in diverse applications. In this study, a unified framework for these partition-based techniques is proposed and the issue of dimension partitions is examined in this framework. An investigation of the relationships of query selectivity and the dimension partition schemes discovers indicators which can predict the performance of a partitioning setting. Accordingly, a greedy algorithm is designed to effectively determine a good partitioning of data dimensions so that the performance of the reduction technique is robust with regard to different datasets. The fourth contribution is an effective similarity search technique in the database of point sets. In the conventional model, an object corresponds to a single vector. In the proposed study, an object is represented by a set of points. In general, this new representation can be used in many real-world applications and carries much more local information, but the retrieval and learning problems become very challenging. The Hausdorff distance is the common distance function to measure the similarity between two point sets, however, this metric is sensitive to outliers in the data. To address this issue, a novel similarity function is defined to better capture the proximity of two objects, in which a one-to-one mapping is established between vectors of the two objects. The optimal mapping minimizes the sum of distances between each paired points. The overall distance of the optimal matching is robust and has high retrieval accuracy. The computation of the new distance function is formulated into the classical assignment problem. The lower-bounding techniques and early-stop mechanism are also proposed to significantly accelerate the expensive similarity search process. The classification problem over the point-set data is called Multiple Instance Learning (MIL) in the machine learning community in which a vector is an instance and an object is a bag of instances. The fifth contribution is to convert the MIL problem into a standard supervised learning in the conventional vector space. Specially, feature vectors of bags are grouped into clusters. Each object is then denoted as a bag of cluster labels, and common patterns of each category are discovered, each of which is further reconstructed into a bag of features. Accordingly, a bag is effectively mapped into a feature space defined by the distances from this bag to all the derived patterns. The standard supervised learning algorithms can be applied to classify objects into pre-defined categories. The results demonstrate that the proposal has better classification accuracy compared to other state-of-the-art techniques. In the future, I will continue to explore my research in large-scale data analysis algorithms, applications and system developments. Especially, I am interested in applications to analyze the massive volume of online data.

Modern machine learning methods are dependent on active optimization research to improve the set of methods available for the efficient and effective extraction of information from large datasets. This, in turn, requires an intense and rigorous study of optimization methods and their possible applications to crucial machine learning applications in order to advance the potential benefits of the field. This thesis provides a study of several modern optimization techniques and supplies a mathematical inquiry into the effectiveness of homotopy methods to attack a fundamental machine learning problem, effective clustering under constraints. The first part of this thesis provides an empirical survey of several popular optimization algorithms, along with one approach that is cutting-edge. These algorithms are tested against deeply challenging real-world problems with vast numbers of local minima, and compares and contrasts the benefits of each when confronted with problems of different natures. The second part of this thesis proposes a new homotopy map for use with constrained clustering problems. This thesis explores the connections between the map and the problem, providing several theorems to justify the use of the map and making use of modern homotopy tracking software to compare an optimization that employs the map with several modern approaches to solving the same problem. / Ph. D.

A l'heure actuelle, les images satellites constituent une source d'information incontournable face à de nombreux enjeux environnementaux (déforestation, caractérisation des paysages, aménagement du territoire, etc.). En raison de leur complexité, de leur volume important et des besoins propres à chaque communauté, l'analyse et l'interprétation des images satellites imposent de nouveaux défis aux méthodes de fouille de données. Le parti-pris de cette thèse est d'explorer de nouvelles approches, que nous situons à mi-chemin entre représentation des connaissances et apprentissage statistique, dans le but de faciliter et d'automatiser l'extraction d'informations pertinentes du contenu de ces images. Nous avons, pour cela, proposé deux nouvelles méthodes qui considèrent les images comme des données quantitatives massives dépourvues de labels sémantiques et qui les traitent en se basant sur les connaissances disponibles. Notre première contribution est une approche hybride, qui exploite conjointement le raisonnement à base d'ontologie et le clustering semi-supervisé. Le raisonnement permet l'étiquetage sémantique des pixels à partir de connaissances issues du domaine concerné. Les labels générés guident ensuite la tâche de clustering, qui permet de découvrir de nouvelles classes tout en enrichissant l'étiquetage initial. Notre deuxième contribution procède de manière inverse. Dans un premier temps, l'approche s'appuie sur un clustering topographique pour résumer les données en entrée et réduire de ce fait le nombre de futures instances à traiter par le raisonnement. Celui-ci n'est alors appliqué que sur les prototypes résultant du clustering, l'étiquetage est ensuite propagé automatiquement à l'ensemble des données de départ. Dans ce cas, l'importance est portée sur l'optimisation du temps de raisonnement et à son passage à l'échelle. Nos deux approches ont été testées et évaluées dans le cadre de la classification et de l'interprétation d'images satellites. Les résultats obtenus sont prometteurs et montrent d'une part, que la qualité de la classification peut être améliorée par une prise en compte automatique des connaissances et que l'implication des experts peut être allégée, et d'autre part, que le recours au clustering topographique en amont permet d'éviter le calcul des inférences sur la totalité des pixels de l'image. / Satellite images have become a valuable source of information for Earth observation. They are used to address and analyze multiple environmental issues such as landscapes characterization, urban planning or biodiversity conservation to cite a few.Despite of the large number of existing knowledge extraction techniques, the complexity of satellite images, their large volume, and the specific needs of each community of practice, give rise to new challenges and require the development of highly efficient approaches.In this thesis, we investigate the potential of intelligent combination of knowledge representation systems with statistical learning. Our goal is to develop novel methods which allow automatic analysis of remote sensing images. We elaborate, in this context, two new approaches that consider the images as unlabeled quantitative data and examine the possible use of the available domain knowledge.Our first contribution is a hybrid approach, that successfully combines ontology-based reasoning and semi-supervised clustering for semantic classification. An inference engine first reasons over the available domain knowledge in order to obtain semantically labeled instances. These instances are then used to generate constraints that will guide and enhance the clustering. In this way, our method allows the improvement of the labeling of existing classes while discovering new ones.Our second contribution focuses on scaling ontology reasoning over large datasets. We propose a two step approach where topological clustering is first applied in order to summarize the data, in term of a set of prototypes, and reduces by this way the number of future instances to be treated by the reasoner. The representative prototypes are then labeled using the ontology and the labels automatically propagated to all the input data.We applied our methods to the real-word problem of satellite images classification and interpretation and the obtained results are very promising. They showed, on the one hand, that the quality of the classification can be improved by automatic knowledge integration and that the involvement of experts can be reduced. On the other hand, the upstream exploitation of topographic clustering avoids the calculation of the inferences on all the pixels of the image.

Depuis la prolifération des bases de données partiellement étiquetées, l'apprentissage automatique a connu un développement important dans le mode semi-supervisé. Cette tendance est due à la difficulté de l'étiquetage des données d'une part et au coût induit de cet étiquetage quand il est possible, d'autre part.L'apprentissage semi-supervisé consiste en général à modéliser une fonction statistique à partir de base de données regroupant à la fois des exemples étiquetés et d'autres non-étiquetés. Pour aborder une telle problématique, deux familles d'approches existent : celles basées sur la propagation de la supervision en vue de la classification supervisée et celles basées sur les contraintes en vue du clustering (non-supervisé). Nous nous intéressons ici à la deuxième famille avec une difficulté particulière. Il s'agit d'apprendre à partir de données avec une partie étiquetée relativement très réduite par rapport à la partie non-étiquetée.Dans cette thèse, nous nous intéressons à l'optimisation des bases de données statistiques en vue de l'amélioration des modèles d'apprentissage. Cette optimisation peut être horizontale et/ou verticale. La première définit la sélection d'instances et la deuxième définit la tâche de la sélection de variables.Les deux taches sont habituellement étudiées de manière indépendante avec une série de travaux considérable dans la littérature. Nous proposons ici de les étudier dans un cadre simultané, ce qui définit la thématique de la co-sélection. Pour ce faire, nous proposons deux cadres unifiés considérant à la fois la partie étiquetée des données et leur partie non-étiquetée. Le premier cadre est basé sur un clustering pondéré sous contraintes et le deuxième sur la préservation de similarités entre les données. Les deux approches consistent à qualifier les instances et les variables pour en sélectionner les plus pertinentes de manière simultanée.Enfin, nous présentons une série d'études empiriques sur des données publiques connues de la littérature pour valider les approches proposées et les comparer avec d'autres approches connues dans la littérature. De plus, une validation expérimentale est fournie sur un problème réel, concernant le diagnostic de transport ferroviaire de l'état de la Thaïlande / We are drowning in massive data but starved for knowledge retrieval. It is well known through the dimensionality tradeoff that more data increase informative but pay a price in computational complexity, which has to be made up in some way. When the labeled sample size is too little to bring sufficient information about the target concept, supervised learning fail with this serious challenge. Unsupervised learning can be an alternative in this problem. However, as these algorithms ignore label information, important hints from labeled data are left out and this will generally downgrades the performance of unsupervised learning algorithms. Using both labeled and unlabeled data is expected to better procedure in semi-supervised learning, which is more adapted for large domain applications when labels are hardly and costly to obtain. In addition, when data are large, feature selection and instance selection are two important dual operations for removing irrelevant information. Both of tasks with semisupervised learning are different challenges for machine learning and data mining communities for data dimensionality reduction and knowledge retrieval. In this thesis, we focus on co-selection of instances and features in the context of semi-supervised learning. In this context, co-selection becomes a more challenging problem as the data contains labeled and unlabeled examples sampled from the same population. To do such semi-supervised coselection, we propose two unified frameworks, which efficiently integrate labeled and unlabeled parts into the co-selection process. The first framework is based on weighting constrained clustering and the second one is based on similarity preserving selection. Both approaches evaluate the usefulness of features and instances in order to select the most relevant ones, simultaneously. Finally, we present a variety of empirical studies over high-dimensional data sets, which are well-known in the literature. The results are promising and prove the efficiency and effectiveness of the proposed approaches. In addition, the developed methods are validated on a real world application, over data provided by the State Railway of Thailand (SRT). The purpose is to propose the application models from our methodological contributions to diagnose the performance of rail dry port systems. First, we present the results of some ensemble methods applied on a first data set, which is fully labeled. Second, we show how can our co-selection approaches improve the performance of learning algorithms over partially labeled data provided by SRT

For the task of near-duplicate document detection, comparison approaches based on bag-of-words used in information retrieval community are not sufficiently accurate. This work presents novel approach when instance-level constraints are given for documents and it is needed to retrieve them, given new query document for near-duplicate detection. The framework incorporates instance-level constraints and clusters documents into groups using novel clustering approach Grouped Latent Dirichlet Allocation (gLDA). Then distance metric is learned for each cluster using large margin nearest neighbor algorithm and finally ranked documents for given new unknown document using learnt distance metrics. The variety of experimental results on various datasets demonstrate that our clustering method (gLDA with side constraints) performs better than other clustering methods and the overall approach outperforms other near-duplicate detection algorithms.

La présente thèse s'intéresse à l'utilisation de méta-données textuelles pour l'analyse d'image. Nous cherchons à utiliser ces informations additionelles comme supervision faible pour l'apprentissage de modèles de reconnaissance visuelle. Nous avons observé un récent et grandissant intérêt pour les méthodes capables d'exploiter ce type de données car celles-ci peuvent potentiellement supprimer le besoin d'annotations manuelles, qui sont coûteuses en temps et en ressources. Nous concentrons nos efforts sur deux types de données visuelles associées à des informations textuelles. Tout d'abord, nous utilisons des images de dépêches qui sont accompagnées de légendes descriptives pour s'attaquer à plusieurs problèmes liés à la reconnaissance de visages. Parmi ces problèmes, la vérification de visages est la tâche consistant à décider si deux images représentent la même personne, et le nommage de visages cherche à associer les visages d'une base de données à leur noms corrects. Ensuite, nous explorons des modèles pour prédire automatiquement les labels pertinents pour des images, un problème connu sous le nom d'annotation automatique d'image. Ces modèles peuvent aussi être utilisés pour effectuer des recherches d'images à partir de mots-clés. Nous étudions enfin un scénario d'apprentissage multimodal semi-supervisé pour la catégorisation d'image. Dans ce cadre de travail, les labels sont supposés présents pour les données d'apprentissage, qu'elles soient manuellement annotées ou non, et absentes des données de test. Nos travaux se basent sur l'observation que la plupart de ces problèmes peuvent être résolus si des mesures de similarité parfaitement adaptées sont utilisées. Nous proposons donc de nouvelles approches qui combinent apprentissage de distance, modèles par plus proches voisins et méthodes par graphes pour apprendre, à partir de données visuelles et textuelles, des similarités visuelles spécifiques à chaque problème. Dans le cas des visages, nos similarités se concentrent sur l'identité des individus tandis que, pour les images, elles concernent des concepts sémantiques plus généraux. Expérimentalement, nos approches obtiennent des performances à l'état de l'art sur plusieurs bases de données complexes. Pour les deux types de données considérés, nous montrons clairement que l'apprentissage bénéficie de l'information textuelle supplémentaire résultant en l'amélioration de la performance des systèmes de reconnaissance visuelle.