http://hdl.handle.net/2117/3689 2013-06-20T05:31:22Z http://hdl.handle.net/2117/19484 Title: Averaging of kernel functions Authors: Belanche Muñoz, Luis Antonio; Tosi, Alessandra Abstract: In kernel-based machines, the integration of several kernels to build more flexible learning methods is a promising avenue for research. In particular, in Multiple Kernel Learning a compound kernel is build by learning a kernel that is the weighted mean of several sources. We show in this paper that the only feasible average for kernel learning is precisely the arithmetic average. We also show that three familiar means (the geometric, inverse root mean square and harmonic means) for positive real values actually generate valid kernels. 2013-06-03T08:59:06Z http://hdl.handle.net/2117/18837 Title: Group-wise sparse correspondences between images based on a common labelling approach Authors: Solé Ribalta, Albert; Sanromà Güell, Gerard; Serratosa Casanelles, Francesc; Alquézar Mancho, René Abstract: Finding sparse correspondences between two images is a usual process needed for several higher-level computer vision tasks. For instance, in robot positioning, it is frequent to make use of images that the robot captures from their cameras to guide the localisation or reduce the intrinsic ambiguity of a specific localisation obtained by other methods. Nevertheless, obtaining good correspondence between two images with a high degree of dissimilarity is a complex task that may lead to important positioning errors. With the aim of increasing the accuracy with respect to the pair-wise image matching approaches, we present a new method to compute group-wise correspondences among a set of images. Thus, pair-wise errors are compensated and better correspondences between images are obtained. These correspondences can be used as a less-noisy input for the localisation process. Group-wise correspondences are computed by finding the common labelling of a set of salient points obtained from the images. Results show a clear increase in effectiveness with respect to methods that use only two images. 2013-04-17T07:41:29Z http://hdl.handle.net/2117/18672 Title: On the computation of the geodesic distance with an application to dimensionality reduction in a neuro-oncology problem Authors: Cruz Barbosa, Raúl; Bautista Villavicencio, David; Vellido Alcacena, Alfredo Abstract: Manifold learning models attempt to parsimoniously describe multivariate data through a low-dimensional manifold embedded in data space. Similarities between points along this manifold are often expressed as Euclidean distances. Previous research has shown that these similarities are better expressed as geodesic distances. Some problems concerning the computation of geodesic distances along the manifold have to do with time and storage restrictions related to the graph representation of the manifold. This paper provides different approaches to the computation of the geodesic distance and the implementation of Dijkstra’s shortest path algorithm, comparing their performances. The optimized procedures are bundled into a software module that is embedded in a dimensionality reduction method, which is applied to MRS data from human brain tumours. The experimental results show that the proposed implementation explains a high proportion of the data variance with a very small number of extracted features, which should ease the medical interpretation of subsequent results obtained from the reduced datasets. 2013-04-08T08:50:58Z http://hdl.handle.net/2117/18670 Title: Comparative diagnostic accuracy of linear and nonlinear feature extraction methods in a neuro-oncology problem Authors: Cruz Barbosa, Raúl; Bautista Villavicencio, David; Vellido Alcacena, Alfredo Abstract: The diagnostic classification of human brain tumours on the basis of magnetic resonance spectra is a non-trivial problem in which dimensionality reduction is almost mandatory. This may take the form of feature selection or feature extraction. In feature extraction using manifold learning models, multivariate data are described through a low-dimensional manifold embedded in data space. Similarities between points along this manifold are best expressed as geodesic distances or their approximations. These approximations can be computationally intensive, and several alternative software implementations have been recently compared in terms of computation times. The current brief paper extends this research to investigate the comparative ability of dimensionality-reduced data descriptions to accurately classify several types of human brain tumours. The results suggest that the way in which the underlying data manifold is constructed in nonlinear dimensionality reduction methods strongly influences the classification results. 2013-04-08T08:37:01Z http://hdl.handle.net/2117/18311 Title: Making machine learning models interpretable Authors: Vellido Alcacena, Alfredo; Martin Guerrero, Jose D.; Lisboa, Paulo J.G. 2013-03-14T16:35:08Z http://hdl.handle.net/2117/18008 Title: Predicting software anomalies using machine learning techniques Authors: Alonso, Javier; Belanche Muñoz, Luis Antonio; Avresky, Dimiter Abstract: In this paper, we present a detailed evaluation of a set of well-known Machine Learning classifiers in front of dynamic and non-deterministic software anomalies. The system state prediction is based on monitoring system metrics. This allows software proactive rejuvenation to be triggered automatically. Random Forest approach achieves validation errors less than 1% in comparison to the well-known ML algorithms under avaluation. In order to reduce automatically the number of monitored parameters, needed to predict software anomalies, we analyze Lasso Regularization technique jointly with the Machine Learning classifiers to evaluate how the prediction accuracy could be guaranteed within an acceptable threshold. This allows to reduce drastically (around 60% in the best case) the number of monitoring parameters. The framework, based on ML and Lasso regularization techniques, has been validated using an e-commerce environment with Apache Tomcat server, and MySql database server. 2013-02-27T15:11:07Z http://hdl.handle.net/2117/18005 Title: A software system for the microbial source tracking problem Authors: Sànchez Mendoza, David; Belanche Muñoz, Luis Antonio; Blanch, Anicet R. Abstract: The aim of this paper is to report the achievement of Ichnaea, a fully computer-based prediction system that is able to make fairly accurate predictions for Microbial Source Tracking studies. The system accepts examples showing diff erent concentration levels, uses indicators (variables) with diff erent environmental persistence, and can be applied at diff erent geographical or climatic areas. We describe the inner workings of the system and report on the specifi c problems and challenges arisen from the machine learning point of view and how they have been addressed. 2013-02-27T12:32:55Z http://hdl.handle.net/2117/18003 Title: Feature selection in proton magnetic resonance spectroscopy data of brain tumors Authors: González Navarro, Félix Fernando; Belanche Muñoz, Luis Antonio Abstract: In cancer diagnosis, classification of the different tumor types is of great importance. An accurate prediction of different tumor types provides better treatment and may minimize the negative impact of incorrectly targeted toxic or aggressive treatments. Moreover, the correct prediction of cancer types using non-invasive information –e.g. 1H-MRS data– could avoid patients to suffer collateral problems derived from exploration techniques that require surgery. A Feature Selection Algorithm specially designed to be use in 1H-MRS Proton Magnetic Resonance Spectroscopy data of brain tumors is presented. It takes advantage of a highly distinctive aspect in this data: some metabolite levels are notoriously different between types of tumors. Experimental read- ings on an international dataset show highly competitive models in terms of accuracy, complexity and medical interpretability. 2013-02-27T11:29:50Z http://hdl.handle.net/2117/18002 Title: Using machine learning techniques to explore 1H-MRS data of brain tumors Authors: González Navarro, Félix Fernando; Belanche Muñoz, Luis Antonio Abstract: Machine learning is a powerful paradigm to analyze Proton Magnetic Resonance Spectroscopy (1H-MRS) spectral data for the classification of brain tumor pathologies. An important characteristic of this task is the high dimensionality of the involved data sets. In this work we apply filter feature selection methods on three types of 1H-MRS spectral data: long echo time, short echo time and an ad hoc combination of both. The experimental findings show that feature selection permits to drastically reduce the dimension, offering at the same time very attractive solutions both in terms of prediction accuracy and the ability to interpret the involved spectral frequencies. A linear dimensionality reduction technique that preserves the class discrimination capabilities is additionally used for visualization of the selected frequencies. 2013-02-27T10:58:24Z http://hdl.handle.net/2117/17558 Title: Leakage localization in water networks using fuzzy logic Authors: Sanz Estapé, Gerard; Pérez Magrané, Ramon; Escobet Canal, Antoni Abstract: This paper presents a methodology for leakage localization using FIR (Fuzzy Inductive Reasoning). A real water network situated in Barcelona has been subdivided in zones which could contain a leakage. Two sensors measure pressures on two separated points of the network. A faulty fuzzy model for each zone and sensor is generated. Test data have been used for classification of leakages in order to evaluate how this methodology helps in leakage localization. Results are compared with another isolation methodology. All the work has been done using simulations carried out by EPANET connected with Matlab. FIR applications used are programmed in Matlab too. 2013-02-01T11:48:26Z http://hdl.handle.net/2117/17482 Title: Cartogram representation of the batch-SOM magnification factor Authors: Tosi, Alessandra; Vellido Alcacena, Alfredo Abstract: Model interpretability is a problem of knowledge extraction from the patterns found in raw data. One key source of knowledge is information visualization, which can help us to gain insights into a problem through graphical representations and metaphors. Nonlinear dimensionality reduction techniques can provide flexible visual insight, but the locally varying representation distortion they produce makes interpretation far from intuitive. In this paper, we define a cartogram method, based on techniques of geographic representation, that allows reintroducing this distortion, measured as a magnification factor, in the visual maps of the batch-SOM model. It does so while preserving the topological continuity of the representation. 2013-01-22T14:55:32Z http://hdl.handle.net/2117/16873 Title: Smooth simultaneous structural graph matching and point-set registration Authors: Sanromà Güell, Gerard; Alquézar Mancho, René; Serratosa Casanelles, Francesc Abstract: We present a graph matching method that encompasses both a model of structural consistency and a model of geometrical deformations. Our method poses the graph matching problem as one of mixture modelling which is solved using the EM algorithm. The solution is then approximated as a succession of assignment problems which are solved, in a smooth way, using Softassign. Our method allows us to detect outliers in both graphs involved in the matching. Unlike the outlier rejectors such as RANSAC and Graph Transformation Matching, our method is able to refine an initial the tentative correspondence-set in a more flexible way than simply removing spurious correspondences. In the experiments, our method shows a good ratio between effectiveness and computational time compared with other methods inside and outside the graphs’ field. 2012-11-09T10:55:23Z http://hdl.handle.net/2117/16224 Title: Remainder subset awareness for feature subset selection Authors: Prat Masramon, Gabriel; Belanche Muñoz, Luis Antonio Abstract: Feature subset selection has become more and more a common topic of research. This popularity is partly due to the growth in the number of features and application domains. It is of the greatest importance to take themost of every evaluation of the inducer, which is normally the more costly part. In this paper, a technique is proposed that takes into account the inducer evaluation both in the current subset and in the remainder subset (its complementary set) and is applicable to any sequential subset selection algorithm at a reasonable overhead in cost. Its feasibility is demonstrated on a series of benchmark data sets. 2012-07-11T09:22:07Z http://hdl.handle.net/2117/16222 Title: A kernel extension to handle missing data Authors: Nebot Troyano, Guillermo; Belanche Muñoz, Luis Antonio Abstract: An extension for univariate kernels that deals with missing values is proposed. These extended kernels are shown to be valid Mercer kernels and can adapt to many types of variables, such as categorical or continuous. The proposed kernels are tested against standard RBF kernels in a variety of benchmark problems showing different amounts of missing values and variable types. Our experimental results are very satisfactory, because they usually yield slight to much better improvements over those achieved with standard methods. 2012-07-11T09:05:34Z http://hdl.handle.net/2117/16220 Title: Machine learning methods for classifying normal vs. tumorous tissue with spectral data Authors: González Navarro, Félix Fernando; Belanche Muñoz, Luis Antonio Abstract: Machine learning is a powerful paradigm within which to analyze 1H-MRS spectral data for the automated classi¯cation of tumor pathologies aimed to facilitate clinical diagnosis. The high dimensionality of the involved data sets makes the discover of computational models a challenging task. In this study we apply a feature selection algorithm in order to reduce the complexity of the problem. The obtained experimental results yield a remarkable classification performance of the final induced models, both in terms of prediction accuracy and number of involved spectral frequencies. A dimensionality reduction technique that preserves the class discrimination capabilities is used for the visualization of the final selected frequencies, thus enhancing their interpretability. 2012-07-10T11:01:34Z