Search

Disorders of consciousness (DoC) happen frequently in various brain injuries. Their detection helps timely treatment for better survival outcomes of DoC patients. It is conventionally undertaken via clinical examinations, typically behavioural assessments. However, these neurological examinations consume significant resources of manpower and time, making continuous DoC monitoring practically infeasible. To address this issue, a computer-aided approach is proposed in this article for automated DoC detection through extracting knowledge from electroencephalogram (EEG) signals. It introduces a new connectivity measure: Power Spectral Density Difference (PSDD) incorporating with a recursive Cosine function (CPSDD). Then, the approach classifies brain-injured patients into DoC (i.e., positive) and wakefulness (i.e., negative) classes through an ensemble of support vector machines (EOSVM), which is a type of machine-learning methods. It is further applied to a dataset of 607 patients with brain injuries. Our classification results show that the EOSVM classifier with the new connectivity measure CPSDD has achieved the best classification performance among 12 connectivity measures. For a setting of 97% majority voting from all SVMs, the EOSVM has diagnosed, in high confidence, 35% of patients with the accuracy, sensitivity, and specificity of 98.21%, 100%, and 95.79%, respectively. Thus, the classifier EOSVM incorporating with the new connectivity measure CPSDD is a promising tool for automatic detection of DoC in brain injuries.

Computer-assisted diagnosis is key for scaling up cervical cancer screening. However, current recognition algorithms perform poorly on whole slide image (WSI) analysis, fail to generalize for diverse staining and imaging, and show sub-optimal clinical-level verification. Here, we develop a progressive lesion cell recognition method combining low- and high-resolution WSIs to recommend lesion cells and a recurrent neural network-based WSI classification model to evaluate the lesion degree of WSIs. We train and validate our WSI analysis system on 3,545 patient-wise WSIs with 79,911 annotations from multiple hospitals and several imaging instruments. On multi-center independent test sets of 1,170 patient-wise WSIs, we achieve 93.5% Specificity and 95.1% Sensitivity for classifying slides, comparing favourably to the average performance of three independent cytopathologists, and obtain 88.5% true positive rate for highlighting the top 10 lesion cells on 447 positive slides. After deployment, our system recognizes a one giga-pixel WSI in about 1.5 min., Computer-assisted diagnosis is key for scaling up cervical cancer screening, but current algorithms perform poorly on whole slide image analysis and generalization. Here, the authors present a WSI classification and top lesion cell recommendation system using deep learning, and achieve comparable results with cytologists.

ObjectiveBrain-computer interface (BCI) can translate intentions directly into instructions and greatly improve the interaction experience for disabled people or some specific interactive applications. To improve the efficiency of BCI, the objective of this study is to explore the feasibility of an audio-assisted visual BCI speller and a deep learning-based single-trial event related potentials (ERP) decoding strategy.ApproachIn this study, a two-stage BCI speller combining the motion-onset visual evoked potential (mVEP) and semantically congruent audio evoked ERP was designed to output the target characters. In the first stage, the different group of characters were presented in the different locations of visual field simultaneously and the stimuli were coded to the mVEP based on a new space division multiple access scheme. And then, the target character can be output based on the audio-assisted mVEP in the second stage. Meanwhile, a spatial-temporal attention-based convolutional neural network (STA-CNN) was proposed to recognize the single-trial ERP components. The CNN can learn 2-dimentional features including the spatial information of different activated channels and time dependence among ERP components. In addition, the STA mechanism can enhance the discriminative event-related features by adaptively learning probability weights.Main resultsThe performance of the proposed two-stage audio-assisted visual BCI paradigm and STA-CNN model was evaluated using the Electroencephalogram (EEG) recorded from 10 subjects. The average classification accuracy of proposed STA-CNN can reach 59.6 and 77.7% for the first and second stages, which were always significantly higher than those of the comparison methods (p < 0.05).SignificanceThe proposed two-stage audio-assisted visual paradigm showed a great potential to be used to BCI speller. Moreover, through the analysis of the attention weights from time sequence and spatial topographies, it was proved that STA-CNN could effectively extract interpretable spatiotemporal EEG features.

Disasters can result in substantial destructive damages to the world. Emergency plan is vital to deal with these disasters. It is still difficult for the traditional CBR to generate emergency plans to meet requirements of rapid responses. An integrated system including Case-based reasoning (CBR) and gravitational search algorithm (GSA) is proposed to generate the disaster emergency plan. Fuzzy GSA (FGSA) is developed to enhance the convergence ability and accomplish the case adaptation in CBR. The proposed algorithm dynamically updates the main parameters of GSA by introducing a fuzzy system. The FGSA-CBR system is proposed, in which fitness function is defined based on the effectiveness of disaster emergency management. The comparison results have revealed that the proposed algorithm has good performances compared with the original GSA and other algorithms. A gas leakage accident is taken as an empirical study. The results have demonstrated that the FGSA-CBR has good performances when generating the disaster emergency plan. The combination of CBR and FGSA can realize the case adaptation, which provides a useful approach to the real applications.

Deep learning models have shown excellent performance in the hyperspectral remote sensing image (HSI) classification. In particular, convolutional neural networks (CNNs) have received widespread attention because of their powerful feature-extraction ability. Recently, a capsule network (CapsNet) was introduced to boost the performance of CNNs, marking a remarkable progress in the field of HSI classification. In this article, we propose a novel deep convolutional capsule neural network (DC-CapsNet) based on spectral–spatial features to improve the performance of CapsNet in the HSI classification while significantly reducing the computation cost of the model. Specifically, a convolutional capsule layer based on the extension of dynamic routing using 3-D convolution is used to reduce the number of parameters and enhance the robustness of the learned spectral–spatial features. Furthermore, a lighter and stronger decoder network composed of deconvolutional layers as a better regularization term and capable of acquiring more spatial relationships is used to further improve the HSI classification accuracy with low computation cost. In this study, we tested the performance of the proposed model on four widely used HSI datasets: the Kennedy Space Center, Indian Pines, Pavia University, and Salinas datasets. We found that the DC-CapsNet achieved high classification accuracy with limited training samples and effectively reduced the computation cost.

Due to the narrow time window for the treatment of acute ischemic stroke, the stroke lesion area in the patient must be identified as quickly and accurately as possible to evaluate the risks and get the most timely and effective treatment. Therefore, there is great clinical significance in the study of automatic identification and segmentation methods for stroke lesions. In this paper, we propose a partitioning-stacking prediction fusion (PSPF) method based on an improved attention U-net to solve the problems of 3D-CNN-based networks, including their high computational cost and insufficient training data, and to achieve accurate segmentation of 3D stroke lesions. Our proposed PSPF method includes three steps in the first part. In Step 1, partitioning, we partition the slices obtained in a certain plane direction by slicing a Magnetic Resonance Imaging (MRI) into subsets according to the 2D graph similarity, then use each partitioned subset to perform training and prediction separately. In Step 2, stacking, we stack the 2D slice results of all subsets according to the position order in MRI before slicing and partitioning to form a 3D lesion result. In Step 3, fusion, we use soft voting to fuse the three orthogonal planes' 3D results that were obtained voxel by voxel in Steps 1 and 2. In the second part, we propose an improved attention U-net, which uses the features from three different scales to generate the attention gating coefficients that further improve training efficiency and segmentation accuracy. We implement a 6-fold cross-validation on the Anatomical Tracings of Lesions After Stroke (ATLAS) dataset to validate our method and model using metrics such as Dice Coefficient (DC), F2 score, and Area under Precision-Recall curve (APR). The results show that compared to the existing methods, our proposed method can not only improve the segmentation precision on unbalanced data but also improve the detailed performance of lesion segmentation. Our proposed method and model are generalized and accurate, demonstrating the good potential for clinical routines. The source codes and models in our method have been made publicly available at [available.upon.acceptance].

Emotion plays a significant role in perceiving external events or situations in daily life. Due to ease of use and relative accuracy, Electroencephalography (EEG)-based emotion recognition has become a hot topic in the affective computing field. However, scalp EEG is a mixed-signal and cannot directly indicate the exact information about active cortex sources of different emotions. In this paper, we analyze the significant differences of active source regions and frequency bands for pairs of emotions-based reconstructed EEG sources using sLORETA, and 26 Brodmann areas are selected as the regions of interest (ROI). And then, six kinds of time- and frequency-domain features from significant active regions and frequency bands are extracted to classify different emotions using support vector machines. Furthermore, we compare the classification performances of emotion features extracted from active source regions and EEG sensors. We have demonstrated that the features from selected source regions can improve the classification accuracy by extensive experiments on the DEAP and TYUT 2.0 EEG-based datasets.

Hyperspectral remote sensing images (HSIs) are rich in spectral–spatial information. The deep learning models can help to automatically extract and discover this rich information from HSIs for classifying HSIs. However, the sampling of the models and the design of the hyperparameters depend on the number of samples and the size of each sample's input space. In the case of limited samples, the description dimension of features is also limited and overfitting to other remote sensing image datasets is evident. This study proposes a novel deep feature aggregation network for HSI classification based on a 3-D convolutional neural network from the perspective of feature aggregation patterns. By introducing the residual learning and dense connectivity strategies, we established a deep feature residual network (DFRN) and a deep feature dense network (DFDN) to exploit the low-, middle-, and high-level features in HSIs. For the Indian Pines and Kennedy Space Center datasets, the DFRN model was determined to be more accurate. On the Pavia University dataset, both the DFDN and DFRN have basically the same accuracy, but the DFDN has faster convergence speed and more stable performance on the validation set than the DFRN. Therefore, when faced with different HSI data, the corresponding aggregation method can be chosen more flexibly according to the requirements on number of training samples and the convergence speed. This is beneficial in the HSI classification.

Land-use information provides a direct representation of the effect of human activities on the environment, and an accurate and efficient land-use classification of remote sensing images is an important element of land-use and land-cover change research. To solve the problems associated with traditional land-use classification methods (e.g., rapid increase in dimensionality of data, inadequate feature extraction, and low running efficiency), a method that combines object-oriented approach with deep convolutional neural network (COCNN) is presented. First, a multi-scale segmentation algorithm is used to segment images to generate image segmentation regions with high homogeneity. Second, a typical rule set of feature objects is constructed on the basis of the object-oriented segmentation results, and the segmentation objects are classified and extracted to form a training sample set. Third, a convolutional neural network (CNN) model structure is modified to improve classification performance, and the training algorithm is optimized to avoid the overfitting phenomenon that occurs during training using small datasets. Ten land-use types are classified by using the remote sensing images covering the area around Fuxian Lake as an example. By comparing the COCNN method with the method based solely on CNN, precision and kappa index were selected to evaluate the classification accuracy of the two methods. For the COCNN method, on the basis of the classification statistics, precision and kappa index coefficients are 96.2% and 0.96, respectively, which are 8.98% and 0.1 higher than those of the method based solely on CNN. Experimental results show that the COCNN method reasonably and efficiently combines object-oriented and deep learning approaches, thereby effectively solving the problem of the inaccurate classification of typical features with better classification accuracy than the simple use of CNN.

Background The ongoing COVID-19 pandemic is characterized by different morbidity and mortality rates across different states, cities, rural areas, and diverse neighborhoods. The absence of a national strategy for battling the pandemic also leaves state and local governments responsible for creating their own response strategies and policies. Objective This study examines the content of COVID-19–related tweets posted by public health agencies in Texas and how content characteristics can predict the level of public engagement. Methods All COVID-19–related tweets (N=7269) posted by Texas public agencies during the first 6 months of 2020 were classified in terms of each tweet’s functions (whether the tweet provides information, promotes action, or builds community), the preventative measures mentioned, and the health beliefs discussed, by using natural language processing. Hierarchical linear regressions were conducted to explore how tweet content predicted public engagement. Results The information function was the most prominent function, followed by the action or community functions. Beliefs regarding susceptibility, severity, and benefits were the most frequently covered health beliefs. Tweets that served the information or action functions were more likely to be retweeted, while tweets that served the action and community functions were more likely to be liked. Tweets that provided susceptibility information resulted in the most public engagement in terms of the number of retweets and likes. Conclusions Public health agencies should continue to use Twitter to disseminate information, promote action, and build communities. They need to improve their strategies for designing social media messages about the benefits of disease prevention behaviors and audiences’ self-efficacy.

For the segmentation task of stroke lesions, using the attention U-Net model based on the self-attention mechanism can suppress irrelevant regions in an input image while highlighting salient features useful for specific tasks. However, when the lesion is small and the lesion contour is blurred, attention U-Net may generate wrong attention coefficient maps, leading to incorrect segmentation results. To cope with this issue, we propose a dual-path attention compensation U-Net (DPAC-UNet) network, which consists of a primary network and auxiliary path network. Both networks are attention U-Net models and identical in structure. The primary path network is the core network that performs accurate lesion segmentation and outputting of the final segmentation result. The auxiliary path network generates auxiliary attention compensation coefficients and sends them to the primary path network to compensate for and correct possible attention coefficient errors. To realize the compensation mechanism of DPAC-UNet, we propose a weighted binary cross-entropy Tversky (WBCE-Tversky) loss to train the primary path network to achieve accurate segmentation and propose another compound loss function called tolerance loss to train the auxiliary path network to generate auxiliary compensation attention coefficient maps with expanded coverage area to perform compensate operations. We conducted segmentation experiments using the 239 MRI scans of the anatomical tracings of lesions after stroke (ATLAS) dataset to evaluate the performance and effectiveness of our method. The experimental results show that the DSC score of the proposed DPAC-UNet network is 6% higher than the single-path attention U-Net. It is also higher than the existing segmentation methods of the related literature. Therefore, our method demonstrates powerful abilities in the application of stroke lesion segmentation.

We propose a novel domain-specific generative pre-training (DSGPT) method for text generation and apply it to the product title and review summarization problems on E-commerce mobile display. First, we adopt a decoder-only transformer architecture, which fits well for fine-tuning tasks by combining input and output all together. Second, we demonstrate utilizing only small amount of pre-training data in related domains is powerful. Pre-training a language model from a general corpus such as Wikipedia or the Common Crawl requires tremendous time and resource commitment, and can be wasteful if the downstream tasks are limited in variety. Our DSGPT is pre-trained on a limited dataset, the Chinese short text summarization dataset (LCSTS). Third, our model does not require product-related human-labeled data. For title summarization task, the state of art explicitly uses additional background knowledge in training and predicting stages. In contrast, our model implicitly captures this knowledge and achieves significant improvement over other methods, after fine-tuning on the public Taobao.com dataset. For review summarization task, we utilize JD.com in-house dataset, and observe similar improvement over standard machine translation methods which lack the flexibility of fine-tuning. Our proposed work can be simply extended to other domains for a wide range of text generation tasks.

Nowadays distributed machine learning (ML) jobs usually adopt a parameter server (PS) framework to train models over large-scale datasets. Such ML job deploys hundreds of concurrent workers, and model parameter updates are exchanged frequently between workers and PSs. Current practice is that workers and PSs may be placed on different physical servers, bringing uncertainty in jobs’ runtime. Also, existing cloud pricing policy often charges a fixed price according to the job’s runtime. Although this pricing strategy is simple to implement, such pricing mechanism is not suitable for distributed ML jobs whose runtime is stochastic and can only be estimated according to its placement after job admission. To supplement existing cloud pricing schemes, we design a dynamic pricing and placement algorithm, DPS, for distributed ML jobs. DPS aims to maximize cloud provider’s profit, which dynamically calculates unit resource price upon a job’s arrival, and determines job’s placement to minimize its runtime if offered price is accepted to users. Our design exploits the multi-armed bandit (MAB) technique to learn unknown information based on past sales. DPS balances the exploration and exploitation stage, and selects the best price based on the reward which is related to job runtime. Our learning-based algorithm increases the provider’s profit, and achieves a sub-linear regret with both the time horizon and the total job number, compared to benchmark pricing schemes. Extensive evaluations also validates the efficacy of DPS.

Electroencephalogram (EEG) analysis has been widely used in the diagnosis of stroke diseases for its low cost and noninvasive characteristics. In order to classify the EEG signals of stroke patients with cerebral infarction and cerebral hemorrhage, this paper proposes a novel EEG stroke signal classification method. This method has two highlights. The first is that a multi-feature fusion method is given by combining wavelet packet energy, fuzzy entropy and hierarchical theory. The second highlight is that a suitable classification model based on ensemble classifier is constructed for perfectly classification stroke signals. Entropy is an accessible way to measure information and uncertainty of time series. Many entropy-based methods have been developed these years. By comparing with the performances of permutation entropy, sample entropy, approximate entropy in measuring the characteristic of stroke patient's EEG signals, it can be found that fuzzy entropy has best performance in characterization stroke EEG signal. By combining hierarchical theory, wavelet packet energy and fuzzy entropy, a multi-feature fusion method is proposed. The method first calculates wavelet packet energy of EEG stroke signal, then extracts hierarchical fuzzy entropy feature by combining hierarchical theory and fuzzy entropy. The experimental results show that, compared with the fuzzy entropy feature, the classification accuracy based on the fusion feature of wavelet packet energy and hierarchical fuzzy entropy is much higher than benchmark methods. It means that the proposed multi-feature fusion method based on stroke EEG signal is an efficient measure in classifying ischemic and hemorrhagic stroke. Support vector machine (SVM), decision tree and random forest are further used as the stroke signal classification models for classifying ischemic stroke and hemorrhagic stroke. Experimental results show that, based on the proposed multi-feature fusion method, the ensemble method of random forest can get the best classification performance in accuracy among three models.

The approaches based on machine learning technology were used to source localization in adverse acoustic environment in recent years. However, the features extracted from the GCC vectors are no significant differences in different location, resulting in an increase in the misclassification. In order to mitigate this problem, this paper introduced the algorithm based on classification of novel cross-correlation function. Firstly, the cross-correlation function is calculated by joint weighting of normalized Phase Transformation and Smooth Coherence Transform (PHAT-SCOT) in each location. Then, acoustic source location is estimated by the Fisher Weighted Naive-Bayes Classifier(FWNBC), which gives different weights according to the importance of features in classification to improve the accuracy. At the same time, the new method was verified in real environment. The experiments all showed that our new method outperforms the baseline algorithm not only in simulating situation but also in real localizing system.

Electroencephalogram (EEG) analysis has been widely used in the diagnosis of stroke diseases for its low cost and noninvasive characteristics. In order to classify the electroencephalogram (EEG) signals of stroke patients with cerebral infarction and cerebral hemorrhage, this paper proposed a novel EEG stroke signal feature extraction method by combining fuzzy entropy and hierarchical theory. Fuzzy entropy not only took the advantages of sample entropy, but also had less dependence on the length of time series and possessed better robustness to noise signals. It measured the similarity of two vectors based on Gaussian function instead of Heaviside function, avoiding discontinuity problems of sample entropy and approximate entropy. Hierarchical theory efficiently took advantages of the approximation information in low-frequency and the detail information in high-frequency. This was benefit for capturing a wealth of dynamic information and retaining redundant components. Support vector machine (SVM) was further used as the stroke signal classification model for classifying ischemic stroke and hemorrhagic stroke. The experimental results showed that, compared with other benchmarks, the classification accuracy based on the features of hierarchical fuzzy entropy is much higher than those benchmarks methods. Compared with the features of fuzzy entropy without using hierarchical theory, the classifier based on the features of hierarchical fuzzy entropy gave a much more improvement in classification performance by increasing accuracy from 68.03% to 96.72%. It meant that the proposed EEG stroke signal hierarchical fuzzy entropy feature extraction method was an efficient measure in classifying ischemic and hemorrhagic stroke.

As a typical cardiovascular disease, stroke is the world's second most deadly disease with the characteristics of high mortality, incidence, recurrence rate and disability rate. To improve the stroke predicting accuracy is a necessary measure for reducing the happening of stroke. As a non-invasive and portable method? Transcranial Doppler (TCD) can provide valuable information in diagnosis of cerebrovascular disease by measuring the blood flow speed in intracranial arteries. How to improve diagnosis accuracy performance based on TCD data set is a necessary demand for realizing cerebrovascular disease TCD intelligent diagnosis system. But TCD data set often presents an imbalanced phenomenon for less patients' data over all examiners. This imbalanced phenomenon easily makes a lower predicting accuracy of cerebrovascular patients. In this paper, an improved fuzzy support vector machine model is proposed by combining class center distance and information entropy. The model efficiently utilizes the advantages of class center distance and information entropy. Moreover, the imbalanced fuzzy support vector machine is used by combining an adaptive imbalanced regulator. The different influence of the imbalance regulating factor value on the classification effect were discussed in detail for finding a suitable regulating factor value. A practical clinical imbalanced TCD dataset is used by collecting clinical TCD data for identify the performance of the proposed method. The experimental results showed that the proposed method can get a better classification performance with higher accuracies and G-means over imbalanced TCD dataset than other benchmark methods.

Ensemble learning is a wildly concerned issue. Traditional ensemble techniques are always adopted to seek better results with labeled data and base classifiers. They fail to address the ensemble task where only unlabeled data are available. A label propagation based ensemble (LPBE) approach is proposed to further combine base classification results with unlabeled data. First, a graph is constructed by taking unlabeled data as vertexes, and the weights in the graph are calculated by correntropy function. Average prediction results are gained from base classifiers, and then propagated under a regularization framework and adaptively enhanced over the graph. The proposed approach is further enriched when small labeled data are available. The proposed algorithms are evaluated on several UCI benchmark data sets. Results of simulations show that the proposed algorithms achieve satisfactory performance compared with existing ensemble methods.

Conventional approaches to acoustic source localization simply based on the received microphone signals, are often vulnerable to adverse acoustic conditions, such as low signal-to-noise ratio (SNR) or high reverberation. But, approaches based on Pattern Recognition and Machine Learning Technology can increase accuracy to locate source in adverse acoustic environment. The advantage of the algorithm is that it requires no calibration of microphone arrays. And Naive Bayes Classifier is simple, fast, and has a small error rate. This paper proposed an improved localization algorithm based on classification of cross-correlation functions (GCC). The weighted cross power spectrum of GCC is smoothed by a smooth filter to formed smooth generalized cross-correlation (SGCC). Then, the classifier model is obtained in each location and form the feature vector. Finally, acoustic source location is estimated by Naive- Bayes classifier. We also proposed in this study the source localization system that based on merely two microphones to input sound signals, combined with improved and optimal methods proposed above. Real-data experiments have demonstrated that algorithm with SGCC has higher localization accuracy than with GCC by 20% in the proposed system at least. The system has good ability to acoustic source localization.

In Geographical Information Systems, geo-coding is used for the task of mapping from implicitly geo-referenced data to explicitly geo-referenced coordinates. At present, an enormous amount of implicitly geo-referenced information is hidden in unstructured text, e.g., Wikipedia, social data and news. Toponym recognition is the foundation of mining this useful geo-referenced information by identifying words as toponyms in text. In this paper, we propose an adapted toponym recognition approach based on deep belief network (DBN) by exploring two key issues: word representation and model interpretation. A Skip-Gram model is used in the word representation process to represent words with contextual information that are ignored by current word representation models. We then determine the core hyper-parameters of the DBN model by illustrating the relationship between the performance and the hyper-parameters, e.g., vector dimensionality, DBN structures and probability thresholds. The experiments evaluate the performance of the Skip-Gram model implemented by the Word2Vec open-source tool, determine stable hyper-parameters and compare our approach with a conditional random field (CRF) based approach. The experimental results show that the DBN model outperforms the CRF model with smaller corpus. When the corpus size is large enough, their statistical metrics become approaching. However, their recognition results express differences and complementarity on different kinds of toponyms. More importantly, combining their results can directly improve the performance of toponym recognition relative to their individual performances. It seems that the scale of the corpus has an obvious effect on the performance of toponym recognition. Generally, there is no adequate tagged corpus on specific toponym recognition tasks, especially in the era of Big Data. In conclusion, we believe that the DBN-based approach is a promising and powerful method to extract geo-referenced information from text in the future.

One of the most popular algorithms for classification is the decision tree. However, existing binary decision tree models do not handle well the minority class over imbalanced data sets. To address this difficulty, a Cost-sensitive and Hybrid attribute measure Multi-Decision Tree (CHMDT) approach is presented in this paper. It penalizes misclassification through a hybrid attribute measure, which is defined from the combination of the Gini index and information gain measure. It further builds a multi-decision tree consisting of multiple decision trees each with different root node information. The overall objective of the approach is to maximize the classification performance with the hybrid attribute measure while minimizing the total misclassification cost. Experiments are conducted over twelve KEEL imbalanced data sets to demonstrate the CHMDT approach. They show that the classification performance of the minority class is improved significantly without sacrifice of the overall classification accuracy of the majority class.

We propose two new SVDD models which improve the robustness to noise.Cutoff distance-based local density can mitigate the effect of noise towards SVDD.Tolerated gap of SVDD with e-insensitive loss can improve generalization performance. As an example of one-class classification methods, support vector data description (SVDD) offers an opportunity to improve the performance of outlier detection and reduce the loss caused by outlier occurrence in many real-world applications. However, due to limited outliers, the SVDD model is built only by using the normal data. In this situation, SVDD may easily lead to over fitting when the normal data contain noise or uncertainty. This paper presents two types of new SVDD methods, named R-SVDD and eNR-SVDD, which are constructed by introducing cutoff distance-based local density of each data sample and the e-insensitive loss function with negative samples. We have demonstrated that the proposed methods can improve the robustness of SVDD for data with noise or uncertainty by extensive experiments on ten UCI datasets. The experimental results have shown that the proposed eNR-SVDD is superior to other existing outlier detection methods in terms of the detection rate and the false alarm rate. Meanwhile, the proposed R-SVDD can also achieve a better outlier detection performance with only normal data. Finally, the proposed methods are successfully used to detect the image-based conveyor belt fault.

Online social networks, such as Facebook, Twitter, and Weibo have played an important role in people's common life. Most existing social network platforms, however, face the challenges of dealing with undesirable users and their malicious spam activities that disseminate content, malware, viruses, etc. to the legitimate users of the service. The spreading of spam degrades user experience and also negatively impacts server-side functions such as data mining, user behavior analysis, and resource recommendation. In this paper, an extreme learning machine (ELM)-based supervised machine is proposed for effective spammer detection. The work first constructs the labeled dataset through crawling Sina Weibo data and manually classifying corresponding users into spammer and non-spammer categories. A set of features is then extracted from message content and user behavior and applies them to the ELM-based spammer classification algorithm. The experiment and evaluation show that the proposed solution provides excellent performance with a true positive rate of spammers and non-spammers reaching 99 and 99.95 %, respectively. As the results suggest, the proposed solution could achieve better reliability and feasibility compared with existing SVM-based approaches.

In this study, a robust classification method for emotional speech single-trial event-related potential (ERP) signal was developed. The classification method based on compression sensing (CS) theory. First, we use CS theory to reduce the dimensionality of the ERP signal. Second, the ERP signal was reconstructed by using K-SVD method to construct the over-complete redundant dictionary. Finally, the ERP signal was classified by calculating the residuals between the reconstructed samples and the test samples. The experimental results show that the proposed algorithm can effectively classify the noisy ERP signal and avoid the feature extraction process in the signal recognition.

In order to extract EEG characteristic waves better, this paper adopts the method of combining wavelet transform with time-frequency blind source separation based on smooth pseudo Wigner-Ville distribution. Firstly, the EEG signal is extracted by wavelet transform to reconstruct the β wave band signal and reconstructed as the initial extracted characteristic wave. Then, to remove the other components which are less relevant to get the enhanced beta wave signal, the time-frequency blind source separation technique based on the smooth pseudo-Wigner distribution is used for the initial extracted Target wave. Finally, the features are extracted, and the support vector machine is used to classify and identify the emotional categories. The experimental results show that the recognition rate is improved when the characteristic wave is extracted by using wavelet transform only.

Selecting an appropriate recognition method is crucial in speech emotion recognition applications. However, the current methods do not consider the relationship between emotions. Thus, in this study, a speech emotion recognition system based on the fuzzy cognitive map (FCM) approach is constructed. Moreover, a new FCM learning algorithm for speech emotion recognition is proposed. This algorithm includes the use of the pleasure-arousal-dominance emotion scale to calculate the weights between emotions and certain mathematical derivations to determine the network structure. The proposed algorithm can handle a large number of concepts, whereas a typical FCM can handle only relatively simple networks (maps). Different acoustic features, including fundamental speech features and a new spectral feature, are extracted to evaluate the performance of the proposed method. Three experiments are conducted in this paper, namely, single feature experiment, feature combination experiment, and comparison between the proposed algorithm and typical networks. All experiments are performed on TYUT2.0 and EMO-DB databases. Results of the feature combination experiments show that the recognition rates of the combination features are 10%–20% better than those of single features. The proposed FCM learning algorithm generates 5%–20% performance improvement compared with traditional classification networks.

Class imbalances have been reported to compromise the performance of most standard classifiers, such as Naive Bayes, Decision Trees and Neural Networks. Aiming to solve this problem, various solutions have been explored mainly via balancing the skewed class distribution or improving the existing classification algorithms. However, these methods pay more attention on the imbalance distribution, ignoring the discriminative ability of features in the context of class imbalance data. In this perspective, a dissimilarity-based method is proposed to deal with the classification of imbalanced data. Our proposed method first removes the useless and redundant features by feature selection from the given data set; and then, extracts representative instances from the reduced data as prototypes; finally, projects the reduced data into a dissimilarity space by constructing new features, and builds the classification model with data in the dissimilarity space. Extensive experiments over 24 benchmark class imbalance data sets show that, compared with seven other imbalance data tackling solutions, our proposed method greatly improves the performance of imbalance learning, and outperforms the other solutions with all given classification algorithms.

As more and more classification algorithms continue to be developed, recommending appropriate algorithms to a given classification problem is increasingly important. This article first distinguishes the algorithm recommendation methods by two dimensions: (1) meta-features, which are a set of measures used to characterize the learning problems, and (2) meta-target, which represents the relative performance of the classification algorithms on the learning problem. In contrast to the existing algorithm recommendation methods whose meta-target is usually in the form of either the ranking of candidate algorithms or a single algorithm, this article proposes a new and natural multilabel form to describe the meta-target. This is due to the fact that there would be multiple algorithms being appropriate for a given problem in practice. Furthermore, a novel multilabel learning-based generic algorithm recommendation method is proposed, which views the algorithm recommendation as a multilabel learning problem and solves the problem by the mature multilabel learning algorithms. To evaluate the proposed multilabel learning-based recommendation method, extensive experiments with 13 well-known classification algorithms, two kinds of meta-targets such as algorithm ranking and single algorithm, and five different kinds of meta-features are conducted on 1,090 benchmark learning problems. The results show the effectiveness of our proposed multilabel learning-based recommendation method.

k-Nearest Neighbor k-NN is one of the most widely used classification algorithms. When classifying a new instance, k-NN first finds out its k nearest neighbors, and then classifies it by voting for the categories of the k nearest neighbors. Therefore, an appropriate number of nearest neighbors is critical for the k-NN classifier. However, in present, there is no systematical solution to determine the specific value of k. In order to address this problem, we propose a novel method of using back-propagation neural networks to explore the relationship between data set characteristics and the optimal values of k, then the relationship and the data set characteristics of a new data set are used to recommend the value of k for this data set. The experimental results on the 49 UCI benchmark data sets show that compared with the optimal k values, although there is a decrease of 1.61% in the average classification accuracy for the k-NN classifier with the recommended k values, the time for determining the k values is greatly shortened.

When the input signals for acoustic echo cancellation (AEC) are related signals, the convergence speed of the traditional normalized least mean square (NLMS) algorithms is significantly reduced. In this paper, a joint optimization robust AEC algorithm is proposed to solve this problem. Based on the analysis of the convergence of the normalized subband adaptive filtering (NSAF) algorithm, the algorithm is optimized by minimizing the mean square error (MSE) of the NSAF algorithm, combining sub-band time-varying step factor and time-varying regularization parameter to update the filter weight vectors. And when the impulse noise occurs, the sub-band cut-off parameter is updated in a recursive manner, which makes the algorithm achieve fast convergence speed and low steady-state error, and has strong robustness to impulse noise. In a series of experiments on AEC, simulation results show that the performance of the algorithm is better than the existing algorithms.

The kernel function is the core of the Support Vector Machine (SVM), and its selection directly affects the performance of SVM. There has been no theoretical basis on choosing a kernel function for speech recognition. In order to improve the learning ability and generalization ability of SVM for speech recognition, this paper presents the Optimal Relaxation Factor (ORF) kernel function, which is a set of new SVM kernel functions for speech recognition, and proves that the ORF function is a Mercer kernel function. The experiments show the ORF kernel function's effectiveness on mapping trend, bi-spiral, and speech recognition problems. The paper draws the conclusion that the ORF kernel function performs better than the Radial Basis Function (RBF), the Exponential Radial Basis Function (ERBF) and the Kernel with Moderate Decreasing (KMOD). Furthermore, the results of speech recognition with the ORF kernel function illustrate higher recognition accuracy.

Sentiment analysis is a technology with great practical value, it can solve the phenomenon of network comment information disorderly to a certain extent, and accurate positioning of user information required. Currently for Chinese sentiment analysis research is relatively small, including a variety of supervised learning method of classification result and the text feature representation methods and feature selection mechanism and other factors impact on the classification performance is an urgent problem. In this paper, we taken the verb, adjectives and adverbs as text features, used TF-IDF to calculate weight of words. Then we adopted the SVM and ELM with kernels to analyze the text emotion tendentiousness. The experimental results show that ELM with kernels can be obtained a better classification result in a relatively short period of time than SVM.

Geo-entity relation recognition from rich texts requires robust and effective solutions on keyword extraction. Compared with supervised learning methods, unsupervised learning methods attract more attention for their capability to capture the dynamic feature variation in text and to discover additional relation types. The frequency-based methods of keyword extraction have been widely studied. However, it is difficult to be applied into geo-entity keyword extraction directly because of the sparse distribution of geo-entity relations in texts. Besides, there are few studies on Chinese keyword extraction. This paper proposes a context enhanced keyword extraction method. Firstly the contexts for geo-entities are enhanced to reduce the sparseness of terms. Secondly two well-known frequency-based statistical methods (i.e., DF and Entropy) are used to build a large-scale corpus automatically from the enhanced contexts. Thirdly the lexical features and their weights are statistically determined based on the corpus to enhance the distinction of the terms. Finally, all terms in the enhanced contexts are measured with the lexical features, and the most important terms are selected as the keywords of geo-entity pairs. Experiments are conducted with mass real Chinese web texts. Compared with DF and Entropy, the presented method improves the precision by 41 % and 36 % respectively in discovering the keywords with sparse distribution and generates additional 60 % correct keywords for geo-entity relation recognition.

Perceptual auditory filter banks such as Bark-scale filter bank are widely used as front-end processing in speech recognition systems. However, the problem of the design of optimized filter banks that provide higher accuracy in recognition tasks is still open. Owing to spectral analysis in feature extraction, an adaptive bands filter bank (ABFB) is presented. The design adopts flexible bandwidths and center frequencies for the frequency responses of the filters and utilizes genetic algorithm (GA) to optimize the design parameters. The optimization process is realized by combining the front-end filter bank with the back-end recognition network in the performance evaluation loop. The deployment of ABFB together with zero-crossing peak amplitude (ZCPA) feature as a front process for radial basis function (RBF) system shows significant improvement in robustness compared with the Bark-scale filter bank. In ABFB, several sub-bands are still more concentrated toward lower frequency but their exact locations are determined by the performance rather than the perceptual criteria. For the ease of optimization, only symmetrical bands are considered here, which still provide satisfactory results.

Aiming at the problem of auditory negative enhancement of typical phase reconstruction method, an improved method of phase reconstruction and MMSE-LSA estimation is proposed. First, the geometric relationship between noisy speech and clean speech in unvoiced segment is used to estimate the phase of the clean speech; Second, considering the randomness of speech appearance in the actual noise environment, a modified MMSE-LSA amplitude estimation is proposed by using the binary hypothesis model. Finally, the new phase reconstruction in voiced and unvoiced speech is combined with the modified MMSE-LSA. The simulation results show that the performance of the algorithm proposed in this paper is better than typical phase reconstruction method in terms of the SegSNR and PESQ.

On the basis of the theory about blind separation of monaural speech based on computational auditory scene analysis (CASA), a two-talker speech separation system combining CASA and speaker recognition was proposed to separate speech from other speech interferences in this paper. First, a tandem algorithm is used to organize voiced speech, then based on the clustering of gammatone frequency cepstral coefficients (GFCCs), an object function is established to recognize the speaker, and the best group is achieved through exhaustive search or beam search, so that voiced speech is organized sequentially. Second, unvoiced segments are generated by estimating onset/offset, and then unvoiced–voiced (U–V) segments and unvoiced–unvoiced (U–U) segments are separated respectively. The U–V segments are managed via the binary mask of the separated voiced speech, while the U–V segments are separated evenly. So far the unvoiced segments are separated. The simulation and performance evaluation verify the feasibility and effectiveness of the proposed algorithm.

Online social networks have played an important role in people's common life. Most existing social network platforms, however, face the challenges of dealing with undesirable users and their malicious spam activities that disseminate content, malware, viruses, etc. to the legitimate users of the service. In this paper, an Extreme Learning Machine based supervised machine is proposed for effective spammer detection. The experiment and evaluation show that the proposed solution provides excellent performance with a true positive rate of spammers and non-spammers reaching 99% and 99.95%, respectively. As the results suggest, the proposed solution could achieve better reliability and feasibility compared with existing SVM based approaches.

Fault detection has become extremely important in industrial production so that numerous potential losses caused from equipment failures could be saved. As a noncontact method, machine vision can satisfy the needs of real-time fault monitoring. However, image-based fault features often have the characteristics of high-dimensionality and redundant correlation. To optimize feature subsets and SVM parameters, this paper presents an enhanced artificial bee colony-based support vector machine (EABC-SVM) approach. The method is applied to the image-based fault detection for the conveyor belt. To improve the optimized capability of original ABC, the EABC algorithm introduces two enhanced strategies including the Cat chaotic mapping initialization and current optimum based search equations. Several UCI datasets have been used to evaluate the performance of EABC-SVM and the experimental results show that this approach has better classification accuracy and convergence performance than the ABC-SVM and other ABC variants-based SVM. Furthermore, the EABC-SVM can achieve a significant detection accuracy of 95% and reduce the amount of features about 65% in the conveyor belt fault detection.

Choosing an appropriate kernel is very important and critical when classifying a new problem with Support Vector Machine. So far, more attention has been paid on constructing new kernels and choosing suitable parameter values for a specific kernel function, but less on kernel selection. Furthermore, most of current kernel selection methods focus on seeking a best kernel with the highest classification accuracy via cross-validation, they are time consuming and ignore the differences among the number of support vectors and the CPU time of SVM with different kernels. Considering the tradeoff between classification success ratio and CPU time, there may be multiple kernel functions performing equally well on the same classification problem. Aiming to automatically select those appropriate kernel functions for a given data set, we propose a multi-label learning based kernel recommendation method built on the data characteristics. For each data set, the meta-knowledge data base is first created by extracting the feature vector of data characteristics and identifying the corresponding applicable kernel set. Then the kernel recommendation model is constructed on the generated meta-knowledge data base with the multi-label classification method. Finally, the appropriate kernel functions are recommended to a new data set by the recommendation model according to the characteristics of the new data set. Extensive experiments over 132 UCI benchmark data sets, with five different types of data set characteristics, eleven typical kernels (Linear, Polynomial, Radial Basis Function, Sigmoidal function, Laplace, Multiquadric, Rational Quadratic, Spherical, Spline, Wave and Circular), and five multi-label classification methods demonstrate that, compared with the existing kernel selection methods and the most widely used RBF kernel function, SVM with the kernel function recommended by our proposed method achieved the highest classification performance.

Blind source separation problem has recently received a great deal of attention in signal processing and unsupervised neural learning. In the current approaches, the additive noise is negligible so that it can be omitted from the consideration. To be applicable in realistic scenarios, blind source separation approaches should deal evenly with the presence of noise. In this contribution, we proposed approaches to independent component analysis when the measured signals are contaminated by additive noise. A noisy multiple channels neural learning algorithm of blind separation is proposed based on independent component analysis. The data have no noise are used to whiten the noisy data, and the windage wipe off technique is used to correct the infection of noise, a neural network model having denoise capability is adopted to recover some original signals from their noisy mixtures observed by the same number of sensors. And a relaxation factor is introduced into the iteration algorithm, thus the new algorithm can implement convergence. Computer simulations and experiment results prove the feasibility and validity of the neural network modeling and control method based on independent component analysis, which can renew the original images effectively.

Many feature subset selection (FSS) algorithms have been proposed, but not all of them are appropriate for a given feature selection problem. At the same time, so far there is rarely a good way to choose appropriate FSS algorithms for the problem at hand. Thus, FSS algorithm automatic recommendation is very important and practically useful. In this paper, a meta learning based FSS algorithm automatic recommendation method is presented. The proposed method first identifies the data sets that are most similar to the one at hand by the k-nearest neighbor classification algorithm, and the distances among these data sets are calculated based on the commonly-used data set characteristics. Then, it ranks all the candidate FSS algorithms according to their performance on these similar data sets, and chooses the algorithms with best performance as the appropriate ones. The performance of the candidate FSS algorithms is evaluated by a multi-criteria metric that takes into account not only the classification accuracy over the selected features, but also the runtime of feature selection and the number of selected features. The proposed recommendation method is extensively tested on 115 real world data sets with 22 well-known and frequently-used different FSS algorithms for five representative classifiers. The results show the effectiveness of our proposed FSS algorithm recommendation method.

In this paper, five different methods in music recognition are discussed, a new character MPCP is proposed in Chord Recognition. The new character overcome the limitation of the traditional PCP and MFCC, apply for recognition system by combining both characteristics. For features we use MPCP vectors, it is trained by CRFs, finally we acquired accurate chords by Viterbi．The experiment show that the proposed strategy can reach a good performance in chord recognition and proved our strategy to be quite promising.

Parameters selection of support vector machine is a very important problem, which has great influence on its performance. In order to improve the learning and generalization ability of support vector machine, in this paper, proposed a new algorithm -parallel artificial fish swarm algorithm to optimize kernel parameter and penalty factor of support vector machine, improved the loop body of artificial fish swarm algorithm to avoid the missing of the optimum solution, and proved its validity by testing with some test functions; used the optimal parameters in a non-specific persons, isolated words, and medium-vocabulary speech recognition system. The experimental results show that the rates of speech recognition based on support vector machine using the new algorithm are better than those of using the traditional artificial fish swarm algorithm in different signal to noise ratios and different words. Especially, the support vector machine model based on the new algorithm can still maintain better recognition rates in lower signal to noise ratios. So the new algorithm is an effective support vector machine parameter optimization method, which makes the support vector machine not only have good generalization ability, but have better robustness.