Word embeddings are widely deployed in a tremendous range of fundamental natural language processing applications and are also useful for generating representations of paragraphs, sentences, and documents. In some contexts involving constrained memory, it may be beneficial to reduce the size of word embeddings since they represent a core component of several natural language processing tasks. By reducing the dimensionality of word embeddings, their usefulness in memory-limited devices can be significantly improved, yielding gains in many real-world applications. This article aims to provide a comparative study of different dimensionality reduction techniques to generate efficient lower-dimensional word vectors. Based on empirical experiments carried out on the Arabic machine translation task, we found that the post-processing algorithm combined with independent component analysis provides optimal performance over the considered dimensionality reduction techniques. Therefore, we arrive at a new combination of the post-processing algorithm and dimensionality reduction (independent component analysis) techniques, which has not been investigated before. The latter was applied to both contextual and non-contextual word embeddings to reduce the size of the vectors while achieving a better translation quality than the original ones.

Feature Selection (FS) is one of the power solutions used in Machine Learning (ML) problems, since it can help to remove irrelevant and redundant attributes, improve the performance, reduce computation time and build more robust models. In this work, a thorough study is carried out to examine the effect of well-performing filter and embedded FS methods for credit scoring. Further, we explore the effect of such methods on the prediction models obtained using different classification techniques. We conduct our experiments on the Australian credit dataset and the obtained experimental results show the benefits of the proposed methodology in credit risk analysis. By considering the selected FS methods with some well chosen classifiers, they make the evaluation more quickly and increase the accuracy of the classification.

Electrocardiogram (ECG) signal is a non-invasive technique that is currently used to diagnose various types of cardiovascular diseases. However, ECG recording is vulnerable to different types of noises and artifacts that make it very difficult to obtain an accurate diagnosis. In this context, we propose a novel ECG denoising algorithm based on the deep Convolutional Denoising Auto-Encoder (CDAE) which requires minimal preprocessing steps, and conserves the important ECG features. In this study, the proposed CDAE algorithm is specifically implemented to remove the Additive White Gaussian noise (AWGN) from the recorded ECG signal. The CDAE was trained, validated and tested on a set of real ECG signals acquired from the well-known MIT-BIH-Arrythmia (MITDB) database with artificially generated AWGN. The experimental results demonstrate that the proposed method shows better Signal to Noise Ratio (SNR) and lower Root Mean Square Error (RMSE) compared to some of the state-of-the-art methods. The promising results indicate also that the proposed CDAE technique is an effective solution for denoising the ECG signal, by providing ECG waves accentuation for other ECG processing applications like diseases diagnosis.

Research on psychological comfort and serenity in social media becomes a necessity because of the excess of negative waves generated by the users. In this paper, we aim to distinguish between offensive and ordinary speech based on machine learning classification techniques. For this purpose, the VAM emotional audio database has been restructured and adjusted to the context of offensive speech detection. Besides, a feature fusion based on Mel Frequency Cepstral Coefficients (MFCCs) as well as Stationary Wavelet Transform (SWT) has been employed and K-nearest neighbors (KNN) algorithm has been used as a classification tool. Results show that the considered feature set has relatively great power in recognizing suspicious behavior, reaching 96.2% as the highest accuracy rate.

Brain-computer interfaces allow us to read human neural data and turn it into valuable information for diverse applications. However, those systems present response disparity because of the non-stationarity of the neural data due to the variability of the response from a subject to another. Thus, they force us to tune the hyperparameters for each case rather than using a unique combination. One method that is used for this purpose is the Bayesian optimization for hyperparameters. In this paper, we propose a study that targets raw EEG signals classified with Convolutional Neural Network and Bayesian Optimization. We optimized the hyperparameters that are related to the architecture. The results suggest it is better to use specific hyperparameters for each subject rather than a global set of hyperparameters for all subjects. Also, we found that Lower Confidence Bound is the best acquisition function for our application.

This paper proposes a novel Deep Learning technique for ECG beats classification. Unlike the traditional Deep Learning models, a new Multi-Scale Wavelet Convolutional Neural Networks (MS-WCNN) is proposed to recognize automatically various cardiac arrhythmias. The proposed MS-WCNN model incorporates the one dimensional CNN and the Stationary Wavelet Transform (SWT) to extract discriminative features from the ECG signal and its wavelet sub-bands simultaneously. The extracted features are then merged using a concatenation strategy. This improves greatly the features learning process of our model at different scales, providing better diagnosis performances. The MITBIH Arrhythmia database has been used to evaluate the performance of the developed model, considering five heartbeats classes: Non-ectopic beat, Supra ventricular ectopic beat, Ventricular ectopic beat, Fusion beat and Unknown beat. The obtained results show that the MS-WCNN method achieves higher or comparable performances with respect to the existing ECG classification algorithms, with an overall diagnosis accuracy of 99, 11%.

In biomedical signal processing, Power Line Interference (50Hz) is one of the most and commonly types of electrical noises that often corrupt the quality of a biomedical data. In this paper, we present a simple tool for ECG signal enhancement approach based on Power Line Interference (PLI) reduction algorithm in Undecimated Wavelet Transform and Interval Thresholding. In our scheme, we use the Undecimated Wavelet Transform (UWT) to overcome the limitations of Discrete Wavelet Transform (DWT) and we present a novel thresholding technique called Interval Thresholding (IT) to overcome the shortcomings of classical thresholding approaches. This method is evaluated on a set of real normal and abnormal ECG recordings from MIT-BIH Arrhythmia database. The results demonstrates that our proposed method gives a better Signal to Noise Ratio improvement (SNRimp) and Higher Cross Correlation coefficient (CC) than four classical ECG denoising approaches such as Notch filter, Wavelet shrinkage, EMD based subtraction and EMD-Wavelet methods.

The latest experiences hint that the QoS (Quality of Service) approaches adopted by IMS (IP Multimedia Subsystem) technologies are still suffering from a primary containment factor due to the nondifferentiation between IPTV (Internet Protocol Television) video components. The success of IPTV services depends on how the customer perceives QoS related to the provided stream. The satisfaction of this factor is crucial to the success of IMS services. This need constitutes a major challenge for the IMS-based IPTV on the horizon of overcoming the failure of existing QoS models namely IntServ and DiffServ. This paper presents a new PHB (Per-Hop Behavior) that reclassifies and differentiates IPTV sub traffics by using the IPv6 Flow Label field. The proposed PHB will make possible prioritization of sub traffic according to the applied QoS network policy. We also suggest the use of OPNET software to implement IMS-Based IPTV scenarios. Results show that the proposed PHB works well and the Video packet losses vary according to the type of IPTV service used by the user.

Several wireless and mobile access technologies such as LTE, 3G, 802.11, WiMaX, etc, are omnipresent in the mobile terminal environment and offer different flows (e.g. data, voice and video). This heterogeneous environment requires the integration of different technologies in order to exploit their complementary characteristics and provide the mobile terminal with the best connectivity anywhere and at anytime, thus respecting the Always Best Connected (ABC) [1] concept. Mobile terminal, equipped with multiple interfaces, can choose the most appropriate access network of another technology among the others that are available in order to improve its quality of service (QoS). Thus, the process of switching between two different technologies it is called Vertical Handover. In this paper, we present a combination between Utility function, Reputation theory and Multi-Attribute Decision Methods for the selection of the best network alternative that ensures a better quality of service requested by the user.

In this paper, we present an alternative approach to the automated aircraft identification system. This system differs from the one proposed in the earlier literature in two ways. First, it uses a general model of the aeronautical air-ground channel which takes into account multipath propagation and Doppler shift. Second, it uses Orthogonal Frequency Division Multiplexing (OFDM) rather than single carrier modulation. The resulting system is then implemented to evaluate the performance of an automatic aircraft identification for the future aeronautical communication system. This automatic identification is achieved by using Spread Spectrum watermarking which allows transmitting embedded digital identification data, such as call sign or tail number, within the speech signal sent by the pilot. We further analyze the impact of different parameters derived from the channel model through simulation results. The system performances are evaluated through transmission reliability and watermark extraction. The experiment results show the impact of channel modeling on the performances of the proposed system.

Variety of vertical handoff algorithms (VHA) based on multi attribute decision making (MADM) methods have been proposed to help the user to select dynamically the best access network in terms of quality of service. However, there is no study that examines the validity of MADM methods which led to inefficient network selection due to inconsistent ranking outcomes. To address this issue, this paper proposes a new validation approach based on group MADM methods. This approach takes into account the weighting algorithms and allows to select the most valid ranking algorithm which can be used in specific traffic classes for network selection decision. Simulation results are presented to illustrate the effectiveness of our new validation approach for the network selection algorithm.

This paper addresses the problem of blind source separation (BSS). To recover original signals, from linear instantaneous mixtures, we propose a new separation technique based on the use of a double referenced system. The reference signals will be incrusted in the cumulant to evaluate the statistical independence between sources. The non-orthogonal joint diagonalization of a set of referenced cumulant matrices gives an estimation of the separating matrix. The important advantage of our proposed technique is the high separation quality in noisy environment by an appropriate choice of reference signals. Computer simulations are presented to illustrate the effectiveness of the suggested approach.

This paper is concerned with blind separation of convolutive mixtures of spatially independent, and temporally possible non linear processes. We consider the MIMO extraction based on the maximization of a contrast function. A new self-styled referenced contrast (RC) function is proposed, which is based on cross-trispectra between the estimated output and a reference signal. Using Parsevals formula, the former criterion yields a new class of time-domain contrast. It presents two main advantages over other more traditional contrasts. Firstly, it concerns the computational cost, and secondly the extension takes into consideration the extraction of the independent sources in the presence of Gaussian sources by making some constraints on the reference signals. There is no comparison with other methods because this is the first technique that deals with this kind of signals in the convolutif mixtures.

This letter addresses the problem of the nonunitary joint block diagonalization of a given set of complex matrices whose potential applications stem from the blind separation of convolutive mixtures of sources and from the array processing. The proposed algorithm is based on the algebraic optimization of a least-mean-square criterion. One of its advantage is that a pre-whitening stage is no more compulsorily required when this algorithm is applied in the blind source separation context. Computer simulations are provided in order to illustrate its behavior in three cases: when exact block-diagonal matrices are built, then when they are progressively perturbed by an additive Gaussian noise and, finally, in the context of blind separation of convolutive mixtures of temporally correlated sources with estimated correlation matrices. A comparison with a classical orthogonal joint block diagonalization algorithm is also performed, and a new performance index is introduced to measure the performance of the separation.

This paper deals with the problem concerning the framework of rotating machines diagnostics by using signal processing advanced tools and more precisely blind source separation (BSS) methods. An application on gear box is given, the objective is to separate gear mesh signals corresponding to each reducer's wheel. It enables us to diagnose and separate each defect in the event of degradation. The proposed method exploits the information redundancy around the meshing frequency and its harmonics resulting from cyclostationarity properties. This redundancy allows us to separate the contribution of each wheel from only one sensor, by tacking advantage of the non-uniformity of the mechanical structure frequency response (MSFR) connecting the exciting source to the sensor

In this letter, we consider contrast functions, which are useful tools for the sources separation problem. We present new generalized contrasts by considering a so-called reference signal. In particular, this allows us to show that a criterion recently proposed in the literature is a contrast. Furthermore, a link with a joint-diagonalization criterion is also emphasized. Finally, we show that another classical contrast can also be extended by considering a reference signal.