A mixture model-based approach to the clustering of microarray expression data

Motivation: This paper introduces the software EMMIX-GENE that has been developed for the specific purpose of a model-based approach to the clustering of microarray expression data, in particular, of tissue samples on a very large number of genes. The latter is a nonstandard problem in parametric cluster analysis because the dimension of the feature space (the number of genes) is typically much greater than the number of tissues. A feasible approach is provided by first selecting a subset of the genes relevant for the clustering of the tissue samples by fitting mixtures of t distributions to rank the genes in order of increasing size of the likelihood ratio statistic for the test of one versus two components in the mixture model. The imposition of a threshold on the likelihood ratio statistic used in conjunction with a threshold on the size of a cluster allows the selection of a relevant set of genes. However, even this reduced set of genes will usually be too large for a normal mixture model to be fitted directly to the tissues, and so the use of mixtures of factor analyzers is exploited to reduce effectively the dimension of the feature space of genes. Results: The usefulness of the EMMIX-GENE approach for the clustering of tissue samples is demonstrated on two well-known data sets on colon and leukaemia tissues. For both data sets, relevant subsets of the genes are able to be selected that reveal interesting clusterings of the tissues that are either consistent with the external classification of the tissues or with background and biological knowledge of these sets.

Text classification using compression-based dissimilarity measures

Arguably, the most difficult task in text classification is to choose an appropriate set of features that allows machine learning algorithms to provide accurate classification. Most state-of-the-art techniques for this task involve careful feature engineering and a pre-processing stage, which may be too expensive in the emerging context of massive collections of electronic texts. In this paper, we propose efficient methods for text classification based on information-theoretic dissimilarity measures, which are used to define dissimilarity-based representations. These methods dispense with any feature design or engineering, by mapping texts into a feature space using universal dissimilarity measures; in this space, classical classifiers (e.g. nearest neighbor or support vector machines) can then be used. The reported experimental evaluation of the proposed methods, on sentiment polarity analysis and authorship attribution problems, reveals that it approximates, sometimes even outperforms previous state-of-the-art techniques, despite being much simpler, in the sense that they do not require any text pre-processing or feature engineering.

Automatic Small Bowel Tumor Diagnosis by Using Multi-Scale Wavelet-Based Analysis in Wireless Capsule Endoscopy Images

BACKGROUND: Wireless capsule endoscopy has been introduced as an innovative, non-invasive diagnostic technique for evaluation of the gastrointestinal tract, reaching places where conventional endoscopy is unable to. However, the output of this technique is an 8 hours video, whose analysis by the expert physician is very time consuming. Thus, a computer assisted diagnosis tool to help the physicians to evaluate CE exams faster and more accurately is an important technical challenge and an excellent economical opportunity. METHOD: The set of features proposed in this paper to code textural information is based on statistical modeling of second order textural measures extracted from co-occurrence matrices. To cope with both joint and marginal non-Gaussianity of second order textural measures, higher order moments are used. These statistical moments are taken from the two-dimensional color-scale feature space, where two different scales are considered. Second and higher order moments of textural measures are computed from the co-occurrence matrices computed from images synthesized by the inverse wavelet transform of the wavelet transform containing only the selected scales for the three color channels. The dimensionality of the data is reduced by using Principal Component Analysis. RESULTS: The proposed textural features are then used as the input of a classifier based on artificial neural networks. Classification performances of 93.1% specificity and 93.9% sensitivity are achieved on real data. These promising results open the path towards a deeper study regarding the applicability of this algorithm in computer aided diagnosis systems to assist physicians in their clinical practice.

Sound Texture Modelling

In the PhD thesis “Sound Texture Modeling” we deal with statistical modelling or textural sounds like water, wind, rain, etc. For synthesis and classification. Our initial model is based on a wavelet tree signal decomposition and the modeling of the resulting sequence by means of a parametric probabilistic model, that can be situated within the family of models trainable via expectation maximization (hidden Markov tree model ). Our model is able to capture key characteristics of the source textures (water, rain, fire, applause, crowd chatter ), and faithfully reproduces some of the sound classes. In terms of a more general taxonomy of natural events proposed by Graver, we worked on models for natural event classification and segmentation. While the event labels comprise physical interactions between materials that do not have textural propierties in their enterity, those segmentation models can help in identifying textural portions of an audio recording useful for analysis and resynthesis. Following our work on concatenative synthesis of musical instruments, we have developed a pattern-based synthesis system, that allows to sonically explore a database of units by means of their representation in a perceptual feature space. Concatenative syntyhesis with “molecules” built from sparse atomic representations also allows capture low-level correlations in perceptual audio features, while facilitating the manipulation of textural sounds based on their physical and perceptual properties. We have approached the problem of sound texture modelling for synthesis from different directions, namely a low-level signal-theoretic point of view through a wavelet transform, and a more high-level point of view driven by perceptual audio features in the concatenative synthesis setting. The developed framework provides unified approach to the high-quality resynthesis of natural texture sounds. Our research is embedded within the Metaverse 1 European project (2008-2011), where our models are contributting as low level building blocks within a semi-automated soundscape generation system.

Effective diagnosis of Alzheimer's disease by means of large margin-based methodology.

BACKGROUND Functional brain images such as Single-Photon Emission Computed Tomography (SPECT) and Positron Emission Tomography (PET) have been widely used to guide the clinicians in the Alzheimer's Disease (AD) diagnosis. However, the subjectivity involved in their evaluation has favoured the development of Computer Aided Diagnosis (CAD) Systems. METHODS It is proposed a novel combination of feature extraction techniques to improve the diagnosis of AD. Firstly, Regions of Interest (ROIs) are selected by means of a t-test carried out on 3D Normalised Mean Square Error (NMSE) features restricted to be located within a predefined brain activation mask. In order to address the small sample-size problem, the dimension of the feature space was further reduced by: Large Margin Nearest Neighbours using a rectangular matrix (LMNN-RECT), Principal Component Analysis (PCA) or Partial Least Squares (PLS) (the two latter also analysed with a LMNN transformation). Regarding the classifiers, kernel Support Vector Machines (SVMs) and LMNN using Euclidean, Mahalanobis and Energy-based metrics were compared. RESULTS Several experiments were conducted in order to evaluate the proposed LMNN-based feature extraction algorithms and its benefits as: i) linear transformation of the PLS or PCA reduced data, ii) feature reduction technique, and iii) classifier (with Euclidean, Mahalanobis or Energy-based methodology). The system was evaluated by means of k-fold cross-validation yielding accuracy, sensitivity and specificity values of 92.78%, 91.07% and 95.12% (for SPECT) and 90.67%, 88% and 93.33% (for PET), respectively, when a NMSE-PLS-LMNN feature extraction method was used in combination with a SVM classifier, thus outperforming recently reported baseline methods. CONCLUSIONS All the proposed methods turned out to be a valid solution for the presented problem. One of the advances is the robustness of the LMNN algorithm that not only provides higher separation rate between the classes but it also makes (in combination with NMSE and PLS) this rate variation more stable. In addition, their generalization ability is another advance since several experiments were performed on two image modalities (SPECT and PET).

Discourse Type Clustering using POS n-gram Profiles and High-Dimensional Embeddings

Abstract: To cluster textual sequence types (discourse types/modes) in French texts, K-means algorithm with high-dimensional embeddings and fuzzy clustering algorithm were applied on clauses whose POS (part-ofspeech) n-gram profiles were previously extracted. Uni-, bi- and trigrams were used on four 19th century French short stories by Maupassant. For high-dimensional embeddings, power transformations on the chi-squared distances between clauses were explored. Preliminary results show that highdimensional embeddings improve the quality of clustering, contrasting the use of bi and trigrams whose performance is disappointing, possibly because of feature space sparsity.

Drift compensation of gas sensor array data by common principal component analysis

A new drift compensation method based on Common Principal Component Analysis (CPCA) is proposed. The drift variance in data is found as the principal components computed by CPCA. This method finds components that are common for all gasses in feature space. The method is compared in classification task with respect to the other approaches published where the drift direction is estimated through a Principal Component Analysis (PCA) of a reference gas. The proposed new method ¿ employing no specific reference gas, but information from all gases ¿has shown the same performance as the traditional approach with the best-fitted reference gas. Results are shown with data lasting 7-months including three gases at different concentrations for an array of 17 polymeric sensors.

Analysis, modelling and visualisation of spatiotemporal data for urban studies

The proportion of population living in or around cites is more important than ever. Urban sprawl and car dependence have taken over the pedestrian-friendly compact city. Environmental problems like air pollution, land waste or noise, and health problems are the result of this still continuing process. The urban planners have to find solutions to these complex problems, and at the same time insure the economic performance of the city and its surroundings. At the same time, an increasing quantity of socio-economic and environmental data is acquired. In order to get a better understanding of the processes and phenomena taking place in the complex urban environment, these data should be analysed. Numerous methods for modelling and simulating such a system exist and are still under development and can be exploited by the urban geographers for improving our understanding of the urban metabolism. Modern and innovative visualisation techniques help in communicating the results of such models and simulations. This thesis covers several methods for analysis, modelling, simulation and visualisation of problems related to urban geography. The analysis of high dimensional socio-economic data using artificial neural network techniques, especially self-organising maps, is showed using two examples at different scales. The problem of spatiotemporal modelling and data representation is treated and some possible solutions are shown. The simulation of urban dynamics and more specifically the traffic due to commuting to work is illustrated using multi-agent micro-simulation techniques. A section on visualisation methods presents cartograms for transforming the geographic space into a feature space, and the distance circle map, a centre-based map representation particularly useful for urban agglomerations. Some issues on the importance of scale in urban analysis and clustering of urban phenomena are exposed. A new approach on how to define urban areas at different scales is developed, and the link with percolation theory established. Fractal statistics, especially the lacunarity measure, and scale laws are used for characterising urban clusters. In a last section, the population evolution is modelled using a model close to the well-established gravity model. The work covers quite a wide range of methods useful in urban geography. Methods should still be developed further and at the same time find their way into the daily work and decision process of urban planners. La part de personnes vivant dans une région urbaine est plus élevé que jamais et continue à croître. L'étalement urbain et la dépendance automobile ont supplanté la ville compacte adaptée aux piétons. La pollution de l'air, le gaspillage du sol, le bruit, et des problèmes de santé pour les habitants en sont la conséquence. Les urbanistes doivent trouver, ensemble avec toute la société, des solutions à ces problèmes complexes. En même temps, il faut assurer la performance économique de la ville et de sa région. Actuellement, une quantité grandissante de données socio-économiques et environnementales est récoltée. Pour mieux comprendre les processus et phénomènes du système complexe "ville", ces données doivent être traitées et analysées. Des nombreuses méthodes pour modéliser et simuler un tel système existent et sont continuellement en développement. Elles peuvent être exploitées par le géographe urbain pour améliorer sa connaissance du métabolisme urbain. Des techniques modernes et innovatrices de visualisation aident dans la communication des résultats de tels modèles et simulations. Cette thèse décrit plusieurs méthodes permettant d'analyser, de modéliser, de simuler et de visualiser des phénomènes urbains. L'analyse de données socio-économiques à très haute dimension à l'aide de réseaux de neurones artificiels, notamment des cartes auto-organisatrices, est montré à travers deux exemples aux échelles différentes. Le problème de modélisation spatio-temporelle et de représentation des données est discuté et quelques ébauches de solutions esquissées. La simulation de la dynamique urbaine, et plus spécifiquement du trafic automobile engendré par les pendulaires est illustrée à l'aide d'une simulation multi-agents. Une section sur les méthodes de visualisation montre des cartes en anamorphoses permettant de transformer l'espace géographique en espace fonctionnel. Un autre type de carte, les cartes circulaires, est présenté. Ce type de carte est particulièrement utile pour les agglomérations urbaines. Quelques questions liées à l'importance de l'échelle dans l'analyse urbaine sont également discutées. Une nouvelle approche pour définir des clusters urbains à des échelles différentes est développée, et le lien avec la théorie de la percolation est établi. Des statistiques fractales, notamment la lacunarité, sont utilisées pour caractériser ces clusters urbains. L'évolution de la population est modélisée à l'aide d'un modèle proche du modèle gravitaire bien connu. Le travail couvre une large panoplie de méthodes utiles en géographie urbaine. Toutefois, il est toujours nécessaire de développer plus loin ces méthodes et en même temps, elles doivent trouver leur chemin dans la vie quotidienne des urbanistes et planificateurs.

Machine learning analysis and modeling of interest rate curves

The present research deals with the review of the analysis and modeling of Swiss franc interest rate curves (IRC) by using unsupervised (SOM, Gaussian Mixtures) and supervised machine (MLP) learning algorithms. IRC are considered as objects embedded into different feature spaces: maturities; maturity-date, parameters of Nelson-Siegel model (NSM). Analysis of NSM parameters and their temporal and clustering structures helps to understand the relevance of model and its potential use for the forecasting. Mapping of IRC in a maturity-date feature space is presented and analyzed for the visualization and forecasting purposes.

Interest rates mapping

The present study deals with the analysis and mapping of Swiss franc interest rates. Interest rates depend on time and maturity, defining term structure of the interest rate curves (IRC). In the present study IRC are considered in a two-dimensional feature space - time and maturity. Exploratory data analysis includes a variety of tools widely used in econophysics and geostatistics. Geostatistical models and machine learning algorithms (multilayer perceptron and Support Vector Machines) were applied to produce interest rate maps. IR maps can be used for the visualisation and pattern perception purposes, to develop and to explore economical hypotheses, to produce dynamic asset-liability simulations and for financial risk assessments. The feasibility of an application of interest rates mapping approach for the IRC forecasting is considered as well. (C) 2008 Elsevier B.V. All rights reserved.

ICA as a preprocessing technique for classification

In this paper we propose the use of the independent component analysis (ICA) [1] technique for improving the classification rate of decision trees and multilayer perceptrons [2], [3]. The use of an ICA for the preprocessing stage, makes the structure of both classifiers simpler, and therefore improves the generalization properties. The hypothesis behind the proposed preprocessing is that an ICA analysis will transform the feature space into a space where the components are independent, and aligned to the axes and therefore will be more adapted to the way that a decision tree is constructed. Also the inference of the weights of a multilayer perceptron will be much easier because the gradient search in the weight space will follow independent trajectories. The result is that classifiers are less complex and on some databases the error rate is lower. This idea is also applicable to regression

Classification of textures seen from different distances and under varying illumination direction

Changes in the angle of illumination incident upon a 3D surface texture can significantly alter its appearance, implying variations in the image texture. These texture variations produce displacements of class members in the feature space, increasing the failure rates of texture classifiers. To avoid this problem, a model-based texture recognition system which classifies textures seen from different distances and under different illumination directions is presented in this paper. The system works on the basis of a surface model obtained by means of 4-source colour photometric stereo, used to generate 2D image textures under different illumination directions. The recognition system combines coocurrence matrices for feature extraction with a Nearest Neighbour classifier. Moreover, the recognition allows one to guess the approximate direction of the illumination used to capture the test image

Generative Part-Based Gabor Object Detector

Object detection is a fundamental task of computer vision that is utilized as a core part in a number of industrial and scientific applications, for example, in robotics, where objects need to be correctly detected and localized prior to being grasped and manipulated. Existing object detectors vary in (i) the amount of supervision they need for training, (ii) the type of a learning method adopted (generative or discriminative) and (iii) the amount of spatial information used in the object model (model-free, using no spatial information in the object model, or model-based, with the explicit spatial model of an object). Although some existing methods report good performance in the detection of certain objects, the results tend to be application specific and no universal method has been found that clearly outperforms all others in all areas. This work proposes a novel generative part-based object detector. The generative learning procedure of the developed method allows learning from positive examples only. The detector is based on finding semantically meaningful parts of the object (i.e. a part detector) that can provide additional information to object location, for example, pose. The object class model, i.e. the appearance of the object parts and their spatial variance, constellation, is explicitly modelled in a fully probabilistic manner. The appearance is based on bio-inspired complex-valued Gabor features that are transformed to part probabilities by an unsupervised Gaussian Mixture Model (GMM). The proposed novel randomized GMM enables learning from only a few training examples. The probabilistic spatial model of the part configurations is constructed with a mixture of 2D Gaussians. The appearance of the parts of the object is learned in an object canonical space that removes geometric variations from the part appearance model. Robustness to pose variations is achieved by object pose quantization, which is more efficient than previously used scale and orientation shifts in the Gabor feature space. Performance of the resulting generative object detector is characterized by high recall with low precision, i.e. the generative detector produces large number of false positive detections. Thus a discriminative classifier is used to prune false positive candidate detections produced by the generative detector improving its precision while keeping high recall. Using only a small number of positive examples, the developed object detector performs comparably to state-of-the-art discriminative methods.

Espaces de timbre générés par des réseaux profonds convolutionnels

Il est avant-tout question, dans ce mémoire, de la modélisation du timbre grâce à des algorithmes d'apprentissage machine. Plus précisément, nous avons essayé de construire un espace de timbre en extrayant des caractéristiques du son à l'aide de machines de Boltzmann convolutionnelles profondes. Nous présentons d'abord un survol de l'apprentissage machine, avec emphase sur les machines de Boltzmann convolutionelles ainsi que les modèles dont elles sont dérivées. Nous présentons aussi un aperçu de la littérature concernant les espaces de timbre, et mettons en évidence quelque-unes de leurs limitations, dont le nombre limité de sons utilisés pour les construire. Pour pallier à ce problème, nous avons mis en place un outil nous permettant de générer des sons à volonté. Le système utilise à sa base des plug-ins qu'on peut combiner et dont on peut changer les paramètres pour créer une gamme virtuellement infinie de sons. Nous l'utilisons pour créer une gigantesque base de donnée de timbres générés aléatoirement constituée de vrais instruments et d'instruments synthétiques. Nous entrainons ensuite les machines de Boltzmann convolutionnelles profondes de façon non-supervisée sur ces timbres, et utilisons l'espace des caractéristiques produites comme espace de timbre. L'espace de timbre ainsi obtenu est meilleur qu'un espace semblable construit à l'aide de MFCC. Il est meilleur dans le sens où la distance entre deux timbres dans cet espace est plus semblable à celle perçue par un humain. Cependant, nous sommes encore loin d'atteindre les mêmes capacités qu'un humain. Nous proposons d'ailleurs quelques pistes d'amélioration pour s'en approcher.

Properties of Support Vector Machines

Support Vector Machines (SVMs) perform pattern recognition between two point classes by finding a decision surface determined by certain points of the training set, termed Support Vectors (SV). This surface, which in some feature space of possibly infinite dimension can be regarded as a hyperplane, is obtained from the solution of a problem of quadratic programming that depends on a regularization parameter. In this paper we study some mathematical properties of support vectors and show that the decision surface can be written as the sum of two orthogonal terms, the first depending only on the margin vectors (which are SVs lying on the margin), the second proportional to the regularization parameter. For almost all values of the parameter, this enables us to predict how the decision surface varies for small parameter changes. In the special but important case of feature space of finite dimension m, we also show that there are at most m+1 margin vectors and observe that m+1 SVs are usually sufficient to fully determine the decision surface. For relatively small m this latter result leads to a consistent reduction of the SV number.