Abstract:Random sample partition (RSP) data model distributedly represents a big data set as a set of RSP data blocks stored on a computing cluster. The RSP data model guarantees that the probability distribution of each data block is statistically consistent to the probability distribution of whole big data set. Thus, each RSP data block is a random sample of big data set and can be used to estimate the statistical properties of big data set or establish the classification and regression models. Based on the RSP data model, the big data analysis can be conducted by analyzing RSP data blocks rather than the whole big data set. This significantly reduces the computational complexity and improves the computing performance of cluster system on big data analysis. In this paper, we firstly present the definition, basic theory and generation method of RSP. Second, we introduce an asymptotic ensemble learning framework called Alpha framework used for big data analysis. Third, we discuss the main big data analysis methods based on the RSP data model and Alpha framework, including data exploration & cleaning, probability density function estimation, supervised subspace learning, semi?supervised ensemble learning, clustering ensemble and outlier detection. Finally, we discuss the innovations and advantages of the RSP data model and Alpha framework in big data analysis by using the divide-and-conquer strategy on random samples.

Abstract:Radar signal sorting is the key technology of electronic countermeasures. Extracting and supplementing other new feature parameters is the effective measures of solving the sorting problem of complex modulation radar signals. In view of the ambiguity function’s unique effect on characterizing signal inherent structure, this paper adopts the improved particle swarm optimization (PSO) algorithm to search the slice of ambiguity function main ridge of the considered signals, and then proposes a feature extraction method which bases on the local difference to extract three local area characteristics, that is the sum of characteristic value, the maximum characteristic value, and the characteristic value distribution entropy. These features can well reflect the local difference of the signal waveform structure. To verify the feasibility and effectiveness of the proposed method, three simulation experiments are designed and the fuzzy C-means algorithm is used to test the clustering performance of the extracted three feature parameters. The experimental results show that, when SNR is not lower than 0 dB, the average clustering accuracy rate of six kinds of the considered signals，i.e., LFM，BFSK，CON，QPSK，M-SEQ and BPSK, is 93.2% and the average accuracy of CON, LFM and BFSK signals achieves to 98.7%. When SNR changes from 0 dB to 20 dB, the average clustering accuracy rate keeps above 80.5%. Meanwhile, the time-effectiveness of the proposed model is better than those compared method. These results illustrate the good performance of the extracted local characteristics.

Abstract:The existing MIMO-OFDM channel estimation method based on compressed sensing uses multiple orthogonal matching pursuit algorithm and its improved algorithm. For such a large-scale data reconstruction algorithm has high computational complexity, storage capacity and other issues, MIMO-OFDM sparse channel estimation method based on the gradient pursuit algorithm is proposed. Gradient pursuit algorithm uses the steepest descent method for the objective function optimal solution, namely calculating the search direction of the objective function and the search step with each iteration, and selecting the optimal solution every atom iterative reconstruction values. As used herein, the estimation performance of gradient pursuit algorithm is compared with the performance of traditional least squares estimation algorithm and orthogonal matching pursuit algorithm. The simulation results show that the gradient pursuit algorithm can guarantee a better estimate and reduce the pilot overhead and the computational complexity. Therefour, the efficiency of reconstruction is improved.

Abstract:In the video image based group activity recognition method, the traditional deep learning methods generally use the conventional（maximum / average）pooling to process the convolutional feature. However, these methods do not consider the importance of the key characters in the group activity which influence the classified result of group behavior. Therefore, we propose an attention based model to detect behavior in group activity videos. In order to identify the group behavior correctly in the video image, this model focuses on the key people in the activity and pools convolutional features dynamically according to the weight of the attention. We conduct extensive experiments on two group behavior datasets, CAD (Collective activity dataset) and CAE (Collective activity extended dataset). The recognition accuracy of our model is better than many existing models using conventional pooling structure.

Abstract:The existing multi-source transfer learning algorithms have very few researches on regression problems, and most of them are symmetric two?class classification problems. This paper presents a weighted multi?source TrAdaBoost regression algorithm, in which the error tolerance coefficient is proposed to solve the problem that the sample weight of the source domain is reduced too quickly, thus the effect of the algorithm is improved. Experiments are performed on the modified Friedman #1 regression problem to verify the effectiveness of the algorithm. The error tolerance coefficient can increase the score by approximately 0.01. In this paper, the proposed algorithm is applied to the industry problems of power communication networks, and the anomaly site (sites with a large number of missing services) detection and true value prediction models are proposed. Moreover, the social network analysis methods are used in the feature engineering, and the importance of the site in the topology is fully considered. Finally, experimental results verify the effectiveness of the algorithm.

Abstract:Machine learning involves some implicit sensitive data that may reveal user’s privacy information when attacked by model attacks such as model queries or model tests. In view of the above problems, this paper proposes a sensitivity data privacy protection Mentoring model PATE-T, which provides a strong privacy guarantee for the training data for machine learning. The method combines multiple Master models trained by disjoint sensitive data sets in a black box manner, relying directly on sensitive training data. Disciple is transfer learning by Master’s collection and cannot directly access Master or basic parameters. Disciple’s data field is different but related to the sensitive training data field. In terms of differential privacy, an attacker can query the Disciple and check its internal work, but it cannot obtain the private information of the training data. Experiments show that the privacy protection model proposed in this paper has reached the balance of privacy/practical accuracy on the MNIST data set and SVHN data set,and the results are superior.

Abstract:A model of compress-and-forward system based on heterogeneous multi-relay network is presented, in which relays compress and encode the received analog signals and transmit them into digital signals, then forward to the destination over Gaussian noise channels with different powers. The destination jointly decodes the received digital signal and yields the estimation of the original signal. Theoretical analysis framework is presented and the expression of the system is derived based on the chief executive officer(CEO) problem of Gaussian source and Shannon channel capacity theory. Subsequently, the power allocation is proposed under the constraint of total system power to maximize the SNR performance. Simulation result verifies that, the SNR performance of the proposed compress-and-forward system outperforms that of the traditional amplify-and-forward system.

Abstract:In the era of big data, large?scale data are often contributed by numerous data sources and used by many data?driven applications. Because of different trustworthiness of sources, different sources often produce data conflicts, making it difficult to determine which information is true. In recent years, truth finding has become a research hotspot by finding the most credibility values from multiple sources. The current truth finding methods usually assume that the entity has only one truth, while in reality, entities may have multiple true values. In this paper, we present an approach for multi?truth finding, which transforms the multi?truth finding into an optimization problem. In so doing, we select the values with the highest credibility as truths of entities. We also propose an asymmetric approach to compute support between values and incorporate influences of similar values to measure value credibility for better truth finding. Experiments on several data sets show that the effectiveness of our algorithm outperform the existing state?of?the?art techniques.

Abstract:Classical test of orthogonality of projected subspaces (TOPS)method is exposed to performance breakdown as the spurious peaks can largely appear in the pseudo spectrum. Here a new direction?of?arrival estimation method for wideband sources using weighted TOPS is introduced. First, it uses the single signal subspace of the reference frequency, where the difference between the smallest signal eigenvalue and the largest noise eigenvalue is the maximum. The proposed weighted TOPS uses the squared TOPS method and test of orthogonality of frequency subspaces (TOFS) method to improved DOA estimation performance, which can obtain the weighted matrix by suitable modification to extend the effectiveness of TOPS. Moreover, the weighted TOPS method uses signal subspace projection instead of null?space projection. Finally, direction?of?arrival estimation for wideband sources are estimated by using trace as performance metric instead of singular value decomposition in TOPS. Comparison of TOPS, squared TOPS and TOFS method, the proposed weighted TOPS can achieve reduction in spurious peaks, improve DOA estimation accuracy, and obtain better computational efficiency and resolution performance for closely spaced sources. Numerical simulations validate the effectiveness of the proposed method.

Abstract:A real?time target tracking method based on compressive sense is investigated. Combining the multiply?features and compressive sense target tracking together, the proposed method introduces a random measurement matrix for detecting in features extraction. Based on the boosting?based frame in tracking, the accuracy of confidence interval estimation is improved by using the identities both of the positive and negative bags of multiply?features. With the proposed method, the chosen target can be tracked on line. Experimental results show that this method can achieve higher efficiency and accuracy in cases such as object moving, pose changing and occlusion. Compared with traditional single feature based target tracking methods, the complementarity of two different features extracted from the proposed method can make the tracking process more robust with a stable and real?time performance.

Abstract:To solve the real?time problem of traditional multi?exposure image fusion and ghost elimination in dynamic scenes, a multi?exposure high dynamic range(HDR) image reconstruction algorithm based on gray scale mapping function modeling is proposed. For low dynamic range(LDR) image sequence of arbitrary size, only visual adaptation S?shaped curves with the same the number of gray?scale need to be fitted, rather than the camera resolution pixels. The HDR can be achieved by fusing the best imaging values directly, which can greatly improve the efficiency of the algorithm fusion and achieve real?time requirements for dynamic scene. The ideal state of multi?exposure image can be achieved by the design of the gray level mapping function. The ghost can be eliminated through moving target detection with difference method. Finally a HDR image reflecting the real scene information and unaffected by ghosts can be achieved.

Abstract:3D object surface reconstruction has important applications in modern clinical medicine, scene modeling and forestry survey and so on. In order to better understand the reconstruction of 3D object surface, this paper first introduces the concept of the Alpha shape of the 3D discrete point set. Based on the analysis of surface reconstruction algorithm using self?adaptive step Alpha?shape is proposed. The value of Alpha is updated dynamically using the kd?tree structure and the average distance of k?nearest neighbors, so that the algorithm can reconstruct the surface with less number of times when the density of the point set is larger. Thus, the reconstruction effect is improved and the operation efficiency of the algorithm is improved. The experimental results with a large number of random data and realistic 3D scanning data show that the proposed algorithm can greatly improve the efficiency compared with the original algorithm.

Abstract:There are a variety of situational elements, element attributes, and intricate relationships between the elements in cyberspace. The deccription of these information directly impact the accuracy, completeness and effectiveness of the cyberspace situation model. The knowledge representation is adopted to describe the key information elements in cyberspace. Firstly, the characteristics of cyberspace situation information knowledge are analyzed, and the important role of knowledge representation in cyberspace situation information is proposed. Secondly, the ontology?based knowledge representation theory is studied, and the motivation of adopting ontology to represent the cyberspace situation is analyzed. Finally, an ontology?based knowledge representation method for cyberspace situational information elements is proposed, and the realization method of that is introduced in detail. The research can effectively promote the perception, modeling and visualization of cyberspace situation, and provide an effective reference for the research of cyberspace related technologies.

Abstract:We consider the two?dimensional (2D) central direction of arrival (DOA) estimation of coherently distributed (CD) source. Making use of the symmetric property of the three?axis crossed array, the rooting method and the generalized ESPRIT algorithm are applied to estimate the central elevation angle and central azimuth angle of the CD source, respectively. The proposed algorithm estimates the 2D DOA independently of deterministic angular distribution function of the distributed source, and requires only one dimensional search. Moreover, the high estimation accuracy is obtained with the help of noncircular signal. Computer simulations show the efficiency of the proposed algorithm.

Abstract:Deep learning has developed rapidly in recent years. The concept of deep learning originates from the neural networks. And the activation function is an indispensable part of the neural network model in learning to understand non-linear functions. Therefore, the common activation functions are studied and compared, aiming at the problems of slow convergence speed, local minimum or gradient disappearance of the commonly used activation functions in back propagation neural networks. In this paper, the Sigmoid and ReLU activation functions are compared, their performances are discussed respectively, and the advantages and disadvantages of several common activation functions are analyzed in detail. Finally, a new activation function, ArcReLU, is proposed by studying the possibility of applying Arctan functions in neural networks and combining with ReLU functions. Experiments show that the function can not only significantly accelerate the training speed of BP neural network, but also effectively reduce the training error and avoid the problem of gradient disappearance.

Abstract:Aiming at the problem of multi-scale pedestrian detection, a pedestrian detection algorithm based on fusion feature pyramid networks (FPN) and faster R-CNN (Faster region convolutional neural network) is proposed. Firstly, FPN and region proposal networks (RPN) are fused. Secondly, FPN and Fast R-CNN are fused. Finally, the pedestrian detection algorithm with fusion FPN and Faster R-CNN is trained and tested on Caltech dataset, KITTI dataset, and ETC dataset, respectively. The mAP (mean Average Precision) of this algorithm reaches 69.72%, 69.76% and 89.74% on Caltech dataset, KITTI dataset, and ETC dataset, respectively. Compared with Faster R-CNN, this algorithm not only improves the pedestrian detection accuracy, but also obtains satisfactory detection effect on the problem of multi-scale pedestrian detection.

Abstract:Fuzzy rough set theory has been paid much attention since it can be used to deal with the uncertainty in the real-valued data or even the mixed data. One of the most important applications of fuzzy rough sets is feature selection, and there have existed many related feature selection methods. However, little attention has been paid on fast feature selection algorithms. Data collected in practice generally include noises or possess some instances with less information. Considering to previously select representative instances from the original data set and perform data mining algorithms on the selected instances set, one may reduce the computation of the algorithms. In view of the advantage of instance selection, the instances are firstly selected based on fuzzy rough sets according to the values of the fuzzy lower approximation of instances in this paper. Then, the evaluation measure of feature selection is constructed by using fuzzy rough set-based information entropy of the selected instances, and the corresponding feature selection algorithm is provided to alleviate the computational complexity. Some numerical experiments are conducted to show the efficiency of the proposed fast algorithm, and the reasonable suggestion of the critical parameter is given to determine the number of the selected instances.

Abstract:Spectral clustering transforms data clustering problem into a graph partitioning problem and classifies data points by finding the optimal sub-graphs. The key to spectral clustering is constructing a suitable similarity matrix, which can truly describe the intrinsic structure of the dataset. However, traditional spectral clustering algorithms adopt Gaussian kernel function to construct the similarity matrix, which results in their sensitivity of selection for scale parameter. In addition, the initial cluster centers need randomly determing at the clustering stage and the clustering performance is not stable. The paper presents an algorithm based on message passing. The algorithm uses a density adaptive similarity measure, which can well describe the relations between data points, and it can obtain high-quality cluster centers through message passing mechanism in affinity propagation (AP) clustering. Moreover, the performance of clustering is optimized by the method. Experiments show that the proposed algorithm can effectively deal with the clustering problem of multi-scale datasets. Its clustering performance is very stable, and the clustering quality is better than traditional spectral clustering algorithm and k-means algorithm.

Abstract:In order to further improve the effect of sidelobe suppression, a synthetic aperture radar image sidelobe suppression method based on wavelet transform and spatial variant apodization is proposed in this paper. 2D wavelet decomposition is carried out for the real and imaginary parts of the complex image after imaging respectively, and the sub-channels obtained after the decomposition are then implemented by the spatial variant apodization sidelobe suppression algorithm. The real part data and imaginary part data are reconstructed by using the sub-channel data after the sidelobe suppression. Then, the reconstructed real part data and the imaginary part data are subjected again to spatial variant apodization sidelobe suppression respectively. Finally, the complex image after wavelet transform spatial variant apodization sidelobe suppression is obtained. Experimental results show that the proposed method can effectively suppress the sidelobe of the image without losing the image resolution. Therefore, this proposed method is an effective sidelobe suppression method for synthetic aperture radar images.

Abstract:By combining the method of questionnaire survey and automatic crawling, a lot of useful tourist information such as users’ personal information, users’ ratings of tourist attractions and other tourism data are obtained. Based on the crawled tourism data, a hierarchical sampling method is applied in turn to generate the “Smart Travel” dataset which contains the important demographic information. Then a user clustering?based collaborative filtering algorithm is implemented to compute the semantic similarity between target user and each clustering center after the users’ ratings of tourist attractions in the “Smart Travel” dataset is preprocessed. Finally, a hybrid recommendation list is generated by absorbing the demographic information obtained by the hierarchical sampling model. Experimental results show that compared with the traditional method, two evaluating indicators like the root mean square error (RMSE) and the mean absolute error (MAE) of the presented algorithm reduce 11.5%—64.9% and 18.8%—47.7%， respectively. Meanwhile, compared with the main baselines, the recommendation precision gets a large improvements as well as the recall rate and better recommendation results are obtained ultimately.