Search

Traditional classification algorithms suppose that the sample distribution among classes is balanced. Yet, such an assumption leads to biased performance over the majority class. This paper proposes a Relevant Information-based UnderSampling (RIUS) approach to select the most relevant examples from the majority class to improve the classification performance for imbalanced data scenarios. RIUS builds on the information-preservation principle that extracts the majority class’s underlying structure with fewer samples. Additionally, we couple our RIUS approach to the well-known Clustering-based Undersampling algorithm (CBUS) to enhance the data representation, and named this RIUS enhancement as CRIUS. Experimental results show that RIUS and CRIUS reveal the data’s relevant structure and reduce the loss of information by selecting the most informative instances.

The adaptive kernel filters are sequential learning algorithms that operate in a particular functional space called a reproducing kernel Hilbert space. However, their performance depends on the selection of two hyper-parameters, i.e., kernel bandwidth and learning-rate. Besides, as these algorithms train the model using a sequence of input vectors, their computation scales with the number of samples. In this work, we propose to address the previous challenges of these sequential learning algorithms. The proposed framework, unlike similar methods, maximizes the correntropy function to optimize the bandwidth and learning-rate parameters. Further, we introduce a sparsification strategy based on dimensionality reduction to remove redundant samples. The framework is tested on both synthetic and real-world data sets, showing convergence to relatively low values of mean-square-error.

Dimensionality reduction aims to preserve, as much as possible, the significant structure of high-dimensional data in the low-dimensional space. This allows removing noise and redundant features, which is useful in exploratory data analysis, classification, and regression tasks. There are two main challenges in dimensionality reduction: (1) how to measure the manifold structures; (2) how to quantify the embedding preservation. On the one hand, previous approaches try to measure the manifold structure using variance, dot product, distance, and similarity preservation. On the other hand, usually, the embedding quality is quantified using divergence-based measures such as Kullback–Leibler and Jensen–Shannon. We propose a dimensionality reduction method that minimizes the mismatch between high and low dimensional spaces. Unlike traditional dimensionality reduction formulations, the proposed approach uses a kernel-based cost function to quantify the embedding quality. Our approach is validated on both synthetic and real-world datasets. In terms of visual inspection and quantitative evaluation of neighborhood preservation, results show that our proposal preserves global data structures in the low-dimensional representation.

We present a network connectivity analysis to assess a relevant vector, valuing the contribution of connection node sets in distinguishing between labeled response stimuli. To this end, the piecewise computation of Phase Locking Index is performed, suggesting a combination procedure to reflect the whole recording span with a single relevance value. Further, we use a supervised, statistical thresholding algorithm to reduce the connectivity matrix dimension, holding the links that mostly differentiate the brain responses to each evoked stimulus. Obtained results in an auditory oddball task show that the developed analysis yields a relevant node set for δ and θ waves that becomes more coherent, connected with improved consistency of performed group-level connectivity graphs.

Generative Stochastic Networks (GSN) for supervised tasks generalize the denoising autoencoders by fixing the deepest layer to the output variables (e.g. class) and estimate the input–output joint distribution as the stationary transition operator of a Markov chain. Because of multi-layer network architectures with stochastic neurons, GSN performance depends on the selected architecture and network training. Aiming to improve such a performance, we introduce a supervised kernel-based learning within a GSN framework. Firstly, the considered network model induces a temporal model working as a data filtering that extracts refined data representations. Then, we use the conventional exhaustive search strategy to fix the hidden layer size. Lastly, we propose a novel supervised layer-wise pre-training that initializes the fine tuning stage of the GSN with more discriminative projection matrices favoring the optimization of the non-convex cost function. Initial matrices are computed by maximizing the centered-kernel alignment (CKA) metric, measuring the affinity between projected samples and labels. We evaluate the proposal performance in comparison with Random, AutoEncoders, and Principal Component Analysis approaches. As a result, CKA-based pre-training approach captures the complex dependencies between parameters, increases the convergence speed in the learning stage, and unravels the data distribution to favor the class discrimination for five widely image collections used in classification tasks of image object recognition.

Medical diagnosis and monitoring benefit from exploiting the advantages of motor imagery (MI) training, which highly depends on the proper interpretation of elicited brain activity responses. Convolutional neural networks (CNN) are increasingly used to improve MI classification performance by employing multi-view extracted features (time, frequency, and spatial). However, deep learning gains knowledge from abundant, complex neural models, resulting in poor interpretability assessments. Here, to enhance the understanding of imagined actions, we develop the relevance analysis of topographic time-frequency representation, preserving an adequate classification performance. Namely, to deal better with the subject variability, a 2D feature combination of continuous wavelet transform and common spatial patterns is extracted to feed a Deep&Wide learning model, assessing the relevance of input multi-view representation that contribute the best to the classifier accuracy. We estimate the feature contribution through the information back-propagated across the hidden layers’ weights before predicting the output label. Evaluation, presented in two databases with bi- and four-label MI tasks, proves that the developed Deep&Wide-based relevance analysis reveals insights about the electrodes, time segments, and frequency bands with relevant MI neural responses, favoring explanation of inter and intra-subject variability because of coordination skills.

The accurate labeling of the volcanic earthquake signals is a crucial task in order to estimate the increase in volcanic activity (among other parameters), which contributes to determine a state of volcanic unrest (as a possible precursor of an eruption). Several automatic classification approaches have been proposed in different computer science areas in order to complement the exhaustive human task of manually labeling volcano-seismic events. However, most of these approaches have been designed under the assumption of stationarity; that is, by discarding the changing nature of the volcanic phenomenon that evolves over time. In this work, an adaptive classification strategy based on incremental learning is proposed, which identifies and learns concept drifts of data streams coming from seismic recordings. The proposed strategy uses a classifier ensemble in order to handle recurrent states and classifies data even while true labels are not yet available. Unlike usual experimental protocols in the state–of–the–art, we carried out experiments with seismic data from Villarrica volcano, considering the chronological order of the data during a long period (more than 4 years of registration) to achieve the detection of drifts. The results show that the proposed classification strategy satisfactorily counteracts the changes, in contrast to standard classifiers trained within a traditional learning scheme. The proposed strategy keeps a stable classification accuracy and achieves a reduction of up to 0.28 of the absolute error in episodes of abrupt changes.

Dimensionality reduction (DR) aims to reveal salient properties of high-dimensional (HD) data in a low-dimensional (LD) representation space. Two elements stipulate success of a DR approach: definition of a notion of pairwise relations in the HD and LD spaces, and measuring the mismatch between these relationships in the HD and LD representations of data. This paper introduces a new DR method, termed Kernel-based entropy dimensionality reduction (KEDR), to measure the embedding quality that is based on stochastic neighborhood preservation, involving a Gram matrix estimation of Renyi's α-entropy. The proposed approach is a data-driven framework for information theoretic learning, based on infinitely divisible matrices. Instead of relying upon regular Renyi's entropies, KEDR also computes the embedding mismatch through a parameterized mixture of divergences, resulting in an improved the preservation of both the local and global data structures. Our approach is validated on both synthetic and real-world datasets and compared to several state-of-the-art algorithms, including the Stochastic Neighbor Embedding-like techniques for which DR approach is a data-driven extension (from the perspective of kernel-based Gram matrices). In terms of visual inspection and quantitative evaluation of neighborhood preservation, the obtained results show that KEDR is competitive and promising DR method. A new dimensionality reduction method is proposed based on a kernel enhancement of stochastic neighborhood preservation.A Gramm matrix estimation of Renyi's alpha-entropy is involved asźthe cost function.The proposed approach is a data-driven framework for information theoretic learning.

Classification of birdsong recordings can be naturally formulated as a multiple instance problem, where bags of instances are represented by either features or dissimilarities. In bioacoustics, bags typically correspond to regions of interest in spectrograms, which are detected after a segmentation stage of the audio recordings. In this paper, we use different dissimilarity measures between bags and explore whether the subsequent application of metric learning/adaptation methods and the construction of dissimilarity spaces allow increasing the classification performance of birdsong recordings. A publicly available bioacoustic data set is used for the experiments. Our results suggest, in the first place, that appropriate dissimilarity measures are those which capture most of the overall differences between bags, such as the modified Hausdorff distance and the mean minimum distance; in the second place, they confirm the benefit from adapting the applied dissimilarity measure as well as the potential further enhancement of the classification performance by building dissimilarity spaces and increasing training set sizes.

Background modeling is widely used in visual surveillance systems aiming to facilitate analysis of real-world video scenes. The goal is to discriminate between pixels from foreground objects and those ones from the background. However, real-world scenarios tend to have time and spatial non-stationary variations, being difficult to reveal the foreground and background entities from video data. Here, we propose a novel adaptive background modeling, termed Object-based Selective Updating with Correntropy (OSUC), to support video-based surveillance systems. Our approach that is developed within an adaptive learning framework unveils existing spatio-temporal pixel relationships, making use of a single Gaussian for the model representation stage. Moreover, we introduce a background updating scheme composed of an updating rule that is based on the stochastic gradient algorithm and Correntropy cost function. As a result, this scheme can extract the temporal statistical pixel distribution, at the same time, dealing with non-stationary pixel value fluctuations that affect the background model. Here, an automatic tuning strategy of the cost function bandwidth parameter is carried out that can handle both Gaussian and non-Gaussian noise environments. Besides, to include pixel spatial relationships in the background modeling processing, we introduce an object-based selective learning rate strategy for enhancing the background modeling accuracy. Particularly, an object motion analysis stage is presented to detect and track foreground entities based on pixel intensities and motion direction attained via optical flow computation. Testing is provided on well-known datasets for discriminating between foreground and background that include stationary and non-stationary behaviors. Achieved results show that the OSUC outperforms, in most of the considered cases, the-state-of-the-art approaches with an affordable computational cost. Therefore, the proposed approach is suitable for supporting real-world video-based surveillance systems. Display Omitted We propose a background modeling method to deal with non-stationary conditions.The object-based updating strategy allows to work with SFO and RFO.Non-Gaussian pixel dynamics are modeled using a Correntropy cost function.Analyzing the regions movement direction avoids tracking background objects.Object-based updating strategies improve the performance for indoor scenarios.

A graph pruning approach, Kernel Alignment based Graph Pruning (KAGP), is proposed.KAGP enhances both the local and global data consistencies.KAGP enhances the clustering performance in most of the cases.KAGP avoids the need of a comprehensive user knowledge about its free parameters.KAGP is a suitable alternative to support spectral clustering algorithms. Display Omitted Detection of data structures in spectral clustering approaches becomes a difficult task when dealing with complex distributions. Moreover, there is a need of a real user prior knowledge about the influence of the free parameters when building the graph. Here, we introduce a graph pruning approach, termed Kernel Alignment based Graph Pruning (KAGP), within a spectral clustering framework that enhances both the local and global data consistencies for a given input similarity. The KAGP allows revealing hidden data structures by finding relevant pair-wise relationships among samples. So, KAGP estimates the loss of information during the pruning process in terms of a kernel alignment-based cost function. Besides, we encode the sample similarities using a compactly supported kernel function that allows obtaining a sparse data representation to support spectral clustering techniques. Attained results shows that KAGP enhances the clustering performance in most of the cases. In addition, KAGP avoids the need for a comprehensive user knowledge regarding the influence of its free parameters.

We introduce STOUT (spatio-temporal unifying tomography), a novel method for the source analysis of electroencephalograpic (EEG) recordings, which is based on a physiologically-motivated source representation. Our method assumes that only a small number of brain sources are active throughout a measurement, where each of the sources exhibits focal (smooth but localized) characteristics in space, time and frequency. This structure is enforced through an expansion of the source current density into appropriate spatio-temporal basis functions in combination with sparsity constraints. This approach combines the main strengths of two existing methods, namely Sparse Basis Field Expansions (Haufe et al., 2011) and Time–Frequency Mixed-Norm Estimates (Gramfort et al., 2013). By adjusting the ratio between two regularization terms, STOUT is capable of trading temporal for spatial reconstruction accuracy and vice versa, depending on the requirements of specific analyses and the provided data. Due to allowing for non-stationary source activations, STOUT is particularly suited for the localization of event-related potentials (ERP) and other evoked brain activity. We demonstrate its performance on simulated ERP data for varying signal-to-noise ratios and numbers of active sources. Our analysis of the generators of visual and auditory evoked N200 potentials reveals that the most active sources originate in the temporal and occipital lobes, in line with the literature on sensory processing.

Training a given learning-based forecasting method to a satisfactory level of performance often requires a large dataset. Indeed, any data-driven methods require having examples that are providing a satisfactory representation of what we wish to model to work properly. This often implies using large datasets to be sure that the phenomenon of interest is properly sampled. However, learning from time series composed of too many samples can also be a problem, given that the computational requirements of the learning algorithms can easily grow following a polynomial complexity according to the training set size. In order to identify representative examples of a dataset, we are proposing a methodology using clustering-based stratification of time series to select a training data subset. The principle for constructing a representative sample set using this method consists in selecting heterogeneous instances picked from all the various clusters composing the dataset. Results obtained show that with a small number of training examples, obtained through the proposed clustering-based stratification, we can preserve the performance and improve the stability of models such as artificial neural networks and support vector regression, while training at a much lower computational cost. We illustrate the methodology through forecasting the one-step ahead Hourly Ontario Energy Price (HOEP).

The use of motor imagery discrimination using feature relevance analysis (MIDFR) is investigated for classification tasks based on electroencephalography (EEG) signals. The method addresses the problem of a direct and automatic finding of the time-varying features influencing the most on distinguishing motor imagery tasks. The method introduces a stochastic relevance stage that is primarily used for properly handling the set of short-time features, which are extracted as to make prominent the nonstationary behavior of the EEG data. Furthermore, since it is widely accepted that the motor imagery information is concentrated in the μ and β neural activity bands, we make use of the empirical mode decomposition together with the common-spatial-patterns mapping. We test two different motor imagery databases, using a soft-margin support vector machine classifier that is validated by a 10-fold cross-validation methodology. Since the proposed MIDFR algorithm better encodes neural activity dynamics, experimental results carried out, which are also contrasted with other state-of-the art approaches, show that the proposed approach allows improving detection of MI classification tasks. Besides, the computed relevance on the EEG channels is in accordance with other clinical findings reported in the literature.

Predicting the localization of a protein has become a useful practice for inferring its function. Most of the reported methods to predict subcellular localizations in Gram-negative bacterial proteins make use of standard protein representations that generally do not take into account the distribution of the amino acids and the structural information of the proteins. Here, we propose a protein representation based on the structural information contained in the pairwise statistical contact potentials. The wavelet transform decodes the information contained in the primary structure of the proteins, allowing the identification of patterns along the proteins, which are used to characterize the subcellular localizations. Then, a support vector machine classifier is trained to categorize them. Cellular compartments like periplasm and extracellular medium are difficult to predict, having a high false negative rate. The wavelet-based method achieves an overall high performance while maintaining a low false negative rate, particularly, on “periplasm” and “extracellular medium”. Our results suggest the proposed protein characterization is a useful alternative to representing and predicting protein sequences over the classical and cutting edge protein depictions.

Authors present an approach based on the transformation of the Cepstral domain on Hidden Markov Model, which is employed for the automatic diagnosis of the Obstructive Sleep Apnea syndrome. The approach includes an Electrocardiogram artefacts removal and R wave detection stage. In addition, the system is modeled by a transformation of the Cepstral domain sequence using Hidden Markov Models (HMM). Final decisions are taken with two different approaches: A Hidden Markov Model and Support Vector Machine classifiers, where the parameterization is based on the transformation of HMM by a kernel. Two public databases have been used for experiments. Firstly, Physionet Apnea-ECG Database for building algorithms, and finally, The St. Vincent's University Hospital/University College Dublin Sleep Apnea Database for testing out with a blind independent dataset. Achieved results were up to 99.23% for Physionet Apnea-ECG Database, and 98.64% for The St. Vincent's Database.

Condition monitoring of mechanical systems is an important topic for industry since it improves machine maintenance and reduce the total associated operational cost. In this sense, vibration analysis is a useful tool for failure prevention in rotating machines, and its main challenge is to perform on-line estimation of dynamic behavior, due to non-stationary operating conditions. To this, estimation of both, amplitude and instantaneous frequency, holding most of process information should be carried out. Nevertheless, approaches for estimating those parameters require to have the shaft speed reference signal, which is not always provided in several industrial applications. In this paper, a novel Order Tracking (OT) scheme of estimation is proposed that is based on the state space model that avoids the shaft speed reference signal. The nonlinear oscillatory model designed as frequency tracker is adapted for estimating the phase and the amplitude of each particular harmonic component. Specifically, nonlinear filtering (namely, the Square-Root Cubature Kalman Filter) is used to estimate the spectral components from the vibration signal. The proposed approach is tested and compared with baseline Vold–Kalman Filtering over four different datasets. The obtained results show that proposed approach is robust and it performs with high accuracy estimation of the order component and the instantaneous frequency under different operating conditions; both allow capturing machine dynamic behavior.

A perceptual measure emulates the human vision for image quality assessment. This paper illustrates the evaluation of Region-of-Interest (ROI) coders using perceptual image quality assessments. The goal of this evaluation is to characterize the coder performance by controlling the ROI quality. Perceptual measures are taken into account for evaluation since they behave as a human-made evaluation. Moreover, a perceptual assessment named Wavelet Quality Index (WQI), is introduced as another image coder evaluator. Proposed assessment aims at emulating the human vision by a weighted linear combination of three wavelet-based perceptual measures. We evaluate the following types of ROI-coders: those preserving the quality of ROI by coarse compression of background (Max-Shift coder), and those balancing the quality between ROI and background (SCM-Shift, and BbB-Shift coders). Using considered assessments for the performance evaluation of coders, results show a variation of evaluation by nature of measurement.

Seismic event recognition is an important task for hazard assessment, eruption prediction and risk mitigation, since it can be used to determine the state of a volcano. Usually, expert technicians read features extracted from the seismogram, such as, cepstral derived coefficients, energy centroids, instant frequency, instant envelop, among others. However, there are few studies about the selection of important features for classifying several types of seismic events, i.e., taking into account the temporal contribution of each considered feature. This paper presents a feature selection strategy based on a relevance measure of time-variant features for seismic event classification. In this research, features are selected as those with the maximal information preserved within the time analysis. Since features selection stage is performed by incremental training, a simple k-nearest neighbor classification rule is used to properly determine the dimension of the final feature set. The employed feature extraction and feature selection methodologies are tested on an isolated event recognition task. The database used to test the methodology is composed of the following classes: volcano-tectonic, long period earthquakes, tremors and hybrid events. Data was recorded at the seismic monitoring stations located at the Nevado-del-Ruiz volcano, Colombia. Using a classifier based on hidden Markov models, accomplished results exhibit a performance improvement from 78% to 88% using the proposed methodology in comparison to the state-of-the-art feature sets.

The determination of relevant acoustic information for the inference of articulators position is an open issue. This paper presents a method to estimate those acoustic features better related to articulators movement. The input feature set is based on time-frequency representation calculated from the speech signal, whose parametrization is achieved using the wavelet-packet transform. The main focus is on measuring the relevant acoustic information, in terms of statistical association, for the inference of articulator positions. The rank correlation Kendall coefficient is used as the relevance measure. Attained statistical association is validated using the @g^2 information measure. The maps of relevant time-frequency features are calculated for the MOCHA-TIMIT database, where the articulatory information is represented by trajectories of specific positions in the vocal tract. Relevant maps are estimated over the whole speech signal as well as on specific phones, for which a given articulator is known to be critical. The usefulness of the relevant maps is tested in an acoustic-to-articulatory mapping system based on gaussian mixture models.

In this work, we analyze video data by learning both the spatial and temporal relationships among frames. For this purpose, the nonlinear dimensionality reduction algorithm, Laplacian Eigenmaps, is improved using a multiple kernel learning framework, and it is assumed that the data can be modeled by means of two different graphs: one considering the spatial information (i.e., the pixel intensity similarities) and the other one based on the frame temporal order. In addition, a formulation for automatic tuning of the required free parameters is stated, which is based on a tradeoff between the contribution of each information source (spatial and temporal). Moreover, we proposed a scheme to compute a common representation in a low-dimensional space for data lying in several manifolds, such as multiple videos of similar behaviors. The proposed algorithm is tested on real-world datasets, and the obtained results allow us to confirm visually the quality of the attained embedding. Accordingly, discussed approach is suitable for data representability when considering cyclic movements.

The computer-assisted analysis of biomedical records has become an essential tool in clinical settings. However, current devices provide a growing amount of data that often exceeds the processing capacity of normal computers. As this amount of information rises, new demands for more efficient data extracting methods appear. This paper addresses the task of data mining in physiological records using a feature selection scheme. An unsupervised method based on relevance analysis is described. This scheme uses a least-squares optimization of the input feature matrix in a single iteration. The output of the algorithm is a feature weighting vector. The performance of the method was assessed using a heartbeat clustering test on real ECG records. The quantitative cluster validity measures yielded a correctly classified heartbeat rate of 98.69% (specificity), 85.88% (sensitivity) and 95.04% (general clustering performance), which is even higher than the performance achieved by other similar ECG clustering studies. The number of features was reduced on average from 100 to 18, and the temporal cost was a 43% lower than in previous ECG clustering schemes.

Highlights? This study presents a methodology for OSA detection based on HRV time series. ? The HRV is characterized by dynamic banked features extracted from its spectrogram. ? The spectral splitting scheme to search frequency boundaries is introduced. ? Attained results show similar performance compared with another outcomes. ? An easier clinical interpretation of HRV-derived parameters is obtained. There is a need for developing simple signal processing algorithms for less costly, reliable and noninvasive Obstructive Sleep Apnoea (OSA) diagnosing. One of the promising directions is to provide the OSA analysis based on the heart rate variability (HRV), which clearly shows a non-stationary behavior. So, a feature extraction approach, being capable of capturing the dynamic heart rate information and suitable for OSA detection, remains an open issue. Grounded on discriminating capability of frequency bands of HRV activity between normal and OSA patients, features can be extracted. However, some HRV normal spectrograms resemble like pathological ones, and vice versa; so, prior to extract the feature set, the energy spatial contribution contained in each sub?band should be clarified. This paper presents a methodology for OSA detection based on a set of short-time feature banked features that are extracted from the spectrogram of the HRV time series. The methodology introduces the spectral splitting scheme, which searches for spectral components with alike stochastic behavior improving the OSA detection accuracy. Two different splitting approaches are considered (heuristic and relevance-based); both of them performing minute-by-minute classification comparable with other outcomes that are reported in literature, but avoiding more complex methods or more computed features. For validation purposes, the methodology is tested on 1-min HRV-segments estimated from 50 Physionet database recordings. Using a parallel combining k-nn classifier, the assessed dynamic feature set reaches as much as 80% value of accuracy, for both considered approaches of spectral splitting. Attained results can be oriented in research focused on finding alternative methods used for less costly and noninvasive OSA diagnosing with the additional benefit of easier clinical interpretation of HRV-derived parameters.

Linear dimensionality reduction (DR) is a widely used technique in pattern recognition to control the dimensionality of input data, but it does neither preserve discriminability nor is capable of discovering nonlinear degrees of freedom present in natural observations. More recently, nonlinear dimensionality reduction (NLDR) algorithms have been developed taking advantage of the fact that data may lie on an embedded nonlinear manifold within an high dimensional feature space. Nevertheless, if the input data is corrupted (noise and outliers), most of nonlinear techniques specially Locally Linear Embedding (LLE) do not produce suitable embedding results. The culprit is the Euclidean distance (cost function in LLE) that does not correctly represent the dissimilarity between objects, increasing the error because of corrupted observations. In this work, the Euclidean distance is replaced by the correntropy induced metric (CIM), which is particularly useful to handle outliers. Moreover, we also extend NLDR to handle manifold divided into separated groups or several manifolds at the same time by employing class label information (CLI), yielding a discriminative representation of data on low dimensional space. Correntropy LLE+CLI approach is tested for visualization and classification on noisy artificial and real-world data sets. The obtained results confirm the capabilities of the discussed approach reducing the negative effects of outliers and noise on the low dimensional space. Besides, it outperforms the other NLDR techniques, in terms of classification accuracy.

Highlights? We propose a new method for automatically computing the number of neighbors in LLE. ? We analyze the global and local properties of the embedding results. ? We study manifolds where the density and the intrinsic dimensionality of the neighborhoods are variable. ? Artificial and real-world datasets were tested. The crux in the locally linear embedding algorithm is the selection of the number of nearest neighbors k. Some previous techniques have been developed for finding this parameter based on embedding quality measures. Nevertheless, they do not achieve suitable results when they are tested on several kind of manifolds. In this work is presented a new method for automatically computing the number of neighbors by means of analyzing global and local properties of the embedding results. Besides, it is also proposed a second strategy for choosing the parameter k, on manifolds where the density and the intrinsic dimensionality of the neighborhoods are changeful. The first proposed technique, called preservation neighborhood error, calculates a unique value of k for the whole manifold. Moreover, the second method, named local neighborhood selection, computes a suitable number of neighbors for each sample point in the manifold. The methodologies were tested on artificial and real-world datasets which allow us to visually confirm the quality of the embedding. According to the results our methods aim to find suitable values of k and appropriated embeddings.

This paper presents new a feature transformation technique applied to improve the screening accuracy for the automatic detection of pathological voices. The statistical transformation is based on Hidden Markov Models, obtaining a transformation and classification stage simultaneously and adjusting the parameters of the model with a criterion that minimizes the classification error. The original feature vectors are built up using classic short-term noise parameters and mel-frequency cepstral coefficients. With respect to conventional approaches found in the literature of automatic detection of pathological voices, the proposed feature space transformation technique demonstrates a significant improvement of the performance with no addition of new features to the original input space. In view of the results, it is expected that this technique could provide good results in other areas such as speaker verification and/or identification.

Locally linear embedding (LLE) is a recent unsupervised learning algorithm for non-linear dimensionality reduction of high dimensional data. One advantage of this algorithm is that just two parameters are needed to be set by user: the number of nearest neighbors and a regularization parameter. The choice of the regularization parameter plays an important role in the embedding results. In this paper, an automated method for choosing this parameter is proposed. Besides, in order to objectively qualify the performance of the embedding results, a new measure of embedding quality is suggested. Our approach is experimentally verified on 9 artificial data sets and 2 real world data sets. Numerical results are compared against two methods previously found in the state of art.

Dimensionality reduction has proved to be a beneficial tool in learning problems. Two of the main advantages provided by dimensionality reduction are interpretation and generalization. Typically, dimensionality reduction is addressed in two separate ways: variable selection and feature extraction. However, in the recent years there has been a growing interest in developing combined schemes such as feature extraction with built-in feature selection. In this paper, we look at dimensionality reduction as a rank-deficient problem that embraces variable selection and feature extraction, simultaneously. From our analysis, we derive a weighting algorithm that is able to select and linearly transform variables by fixing the dimensionality of the space where a relevance criterion is evaluated. This step enforces sparseness on the resulting weights. Our main goal is dimensionality reduction for classification problems. Namely, we introduce modified versions of principal component analysis (PCA) by expectation maximization (EM) and linear regularized discriminant analysis (RDA). Finally, we propose a simple extension of WRDA that deals with functional features. In this case, observations are described by a set of functions defined over the same domain. Methods were put to test on artificial and real data sets showing high levels of generalization even for small sized training samples.

Even though, under representational restrictions, the nearest feature rules and the dissimilarity-based classifiers are feasible alternatives to the nearest neighbor method; individually, they may not be sufficiently powerful if a very small set of prototypes is required, e.g. when it is computationally expensive to deal with larger sets of prototypes. In this paper, we show that combining both strategies, taking advantage of their individual properties, provides an improvement, particularly for correlated data sets. The combined strategy consists in deriving an enriched (generalized) dissimilarity representation by using the nearest feature rules, namely feature lines and feature planes. On top of that enriched representation, Bayesian classifiers can be constructed in order to obtain a good generalization.