ISSN 0253-2778

CN 34-1054/N

2020 Vol. 50, No. 8

Display Method:
Original Paper
Influence of firm crisis self-disclosure on stock reaction: Empirical research based on 220 safety accident announcements of Chinese A-share listed firms
WEI Wenzhe, ZHOU Lei, WEI Jiuchang
2020, 50(8): 1035-1047. doi: 10.3969/j.issn.0253-2778.2020.08.001
Based on signal theory and situational crisis communication theory, the abnormal return rate and cumulative abnormal return rate of stocks were calculated by event study method, and the negative reaction process of stocks before and after firm crisis accident was verified. Here 220 safety accident announcements of Chinese A-share listed firms from 2007 to 2018 were collected and quantified as samples. In addition, based on stakeholders’ perception of environmental uncertainty, the multi-regression method was used to explore the process of stakeholders’ response to the self-disclosure of firm crisis. The empirical results show that within a short period after the announcement, stakeholders are more concerned about the loss of personnel, property damage and the response speed of firms. But after a long time, the negative stock reaction to the future uncertainty is stronger. Therefore, the sequence of stakeholders’ response to the self-disclosure of firm crisis is from influence uncertainty, response uncertainty to state uncertainty. The conclusion helps firms to make better use of crisis response strategies to reduce the damage and improve the ability of crisis communication.
Research and practice of plagiarism detection in program code assignments by college students
YU Jun, LI Yajie, CHENG Lilei, LIAN Shun, TAN Chang, DING Decheng, Liu Qi
2020, 50(8): 1048-1057. doi: 10.3969/j.issn.0253-2778.2020.08.002
The programming ability of students directly reflects the learning effect of technical courses. The proportion of program code assignments are increasing in teaching evaluation. The low cost of plagiarism of program code homework leads to the widespread plagiarism in colleges and universities, which seriously affects the cultivation of students’ ability and the effect of teaching. To this end, a method for homework plagiarism detection is proposed by combining the artificial intelligence algorithm with data processing analysis technology to detect similarities in students’ homework intelligently and automatically, and analyze the overall situation of plagiarism. First, the complex situation of the program code assignments submitted by students is analyzed, and the data pre-processing process is designed. Then, the similarity detection algorithm for program code assignments based on KR and Winnowing is specifically proposed. Compared with the traditional detection methods, the accuracy of similarity detection in students’ homework is improved by such means as code formatting. In the practice of large-scale homework detection, the research optimization algorithm increases the differentiation of similarity results in different students’ homework. To verify the validity and practicability of the core similarity calculation part of this paper, a relevant simulation experiment process (including the comparison with JPlag detection system), was designed and the similarity calculation results were given under different plagiarism types on the same experimental data set. Finally, based on iFLYTEK’s Bosi intelligent online learning platform, the research has been applied in real scenarios. The experimental results and practical application results show that the proposed detection method has high validity and application value in the detection of similarities in program code assignments by college students.
Citation network’s influence maximization algorithm based on global influence
ZHANG Wenjing, BAN Zhijie
2020, 50(8): 1058-1063. doi: 10.3969/j.issn.0253-2778.2020.08.003
It is of great significance for academic researches to search out the most influential papers from a huge number of Journal papers. However, the existing algorithms for maximizing influence need to be combined with greedy algorithm, which increases the time complexity. According to the time unidirectional and acyclic features of the citation relationship in the citation network, an algorithm is proposed to maximize the influence based on the global influence of nodes. The algorithm mainly includes: ①Calculating the global influence of all nodes. Combined with the publication time characteristics of the citation network, the upper triangular sparse influence matrix is constructed. On the basis of the linear threshold propagation model, the direct and indirect path effects between nodes and the cumulative calculation rule are used to simulate the propagation process of influence on the network. Every time the square matrix is calculated, the influence of all nodes will be propagated down one hop to get the influence of the next path, and all the influences will be counted to finally get the square matrix representing the global influence of all nodes; ②All nodes will be ranked according to the global influence, and the first n nodes will be selected as candidate nodes to select k seed nodes. By the cumulative calculation rule, the proposed algorithm avoids the overlapping of influence among nodes during the process of selecting seed nodes. The real academic citation network data set is taken as the experimental sample, and our algorithm is compared with the two benchmark algorithms in terms of activation range and running time. Experimental results show that the proposed algorithm greatly reduces the time complexity, and that the activation range is close to the greedy algorithm.
Differential privacy protection method for deep learning based on WGAN feedback
TAO Tao, BAI Jianshu, LIU Heng, HOU Shudong, ZHENG Xiao
2020, 50(8): 1064-1071. doi: 10.3969/j.issn.0253-2778.2020.08.004
Aiming at the problem that attackers may steal sensitive information of the deep learning training dataset by some technological means such as the Generative Adversarial Network(GAN), combining the differential privacy theory, the differential privacy protection method was proposed for deep learning based on the Wasserstein generative adversarial network(WGAN) feedback parameter tuning. This privacy protection method is realized by optimization of the stochastic gradient descent, gradient clipping of setting gradient threshold, and noise adding to the optimization process of deep learning; WGAN was used to generate optimized results similar to the original data. The difference of the generated results and the original data were used for feedback parameter tuning. The experiment result shows that this method can effectively protect sensitive private information of the dataset and has preferable data usability.
Campus encyclopedia platform based on knowledge recommendation
REN Min, XU Ling, WANG Feng, WU Chao
2020, 50(8): 1072-1076. doi: 10.3969/j.issn.0253-2778.2020.08.005
In 2018, University of Science and Technology of China (USTC) released its encyclopedia platform of the university, which provides intelligent and digital means to accumulate and disperse cultural knowledge of the university. The release of Encyclopedia Platform aims to build up a knowledge base of the campus culture, encouraging the staff and students to actively participate in the recording and sharing of the unique campus culture of USTC, and provides intelligent retrieval and knowledge recommendation services. We present a thorough analysis of the key intelligent technologies of the encyclopedia platform, i.e., hierarchical multi-label classification, intelligent retrieval technologies, tags and collaborative filtering based recommendations, and introduce the architecture and the key functions of the platform. Finally, we briefly discuss its usage statistics.
2D HSQC NMR analysis of plant cell wall material in DMSO-d6/HMPA-d18
WANG Wanwan, CAI Jibao, XU Zhenyu, NIU Fanchao, SU Jiakun, ZHANG Yi, YANG Jun
2020, 50(8): 1077-1083. doi: 10.3969/j.issn.0253-2778.2020.08.006
The cell wall material gel is the product of the cell wall material directly formed with deuterated solvent in the nuclear magnetic resonance tube. The well-resolved/dispersed 2D 13C-1H related nuclear magnetic resonance spectra (2D HSQC NMR)can be obtained without the separation of components. Deuterated dimethyl sulfoxide (DMSO-d6), deuterated dimethyl sulfoxide/deuterated pyridine (DMSO-d6/pyridine-d5) and deuterated dimethyl sulfoxide/deuterated hexamethylphosphoryltriamine (DMSO-d6/HMPA-d18) were selected to dissolve the cell wall of poplar (angiosperm). The spectrum signals under DMSO-d6/HMPA-d18 were the most abundant, and the correlation of p-hydroxyphenyl signals could be easily obtained. Therefore, DMSO-d6/ HMPA-d18 improves the resolution and intensity of the spectrum. DMSO-d6/HMPA-d18 was used to characterize the two-dimensional NMR structure of pine (gymnosperms), and the spectrum showed high-resolution polysaccharide correlation and lignin structure. Studies have shown that this method is suitable for the detection of cell wall substances in plants.Therefore, the 2D HSQC NMR study under DMSO-d6/HMPA-d18 is a faster and greener method to evaluate the structure of plant cell wall.
Design of quad-rotor general controller based on ensemble modeling method
WU Shichong, LIAO Fei, WU Wenhua, FU Zaiming
2020, 50(8): 1084-1092. doi: 10.3969/j.issn.0253-2778.2020.08.007
To solve the problem of quad-rotor UAV general controller approximation, a method is proposed based on ensemble modeling to learn and construct the general form of controller. Quadrotor hover and forward flight tasks are designed and simulated on Matlab/Simulink to obtain training and test data sets. Then the state variables and process variables of the quadrotor flight are taken as inputs, and the lift forces of the rotors are taken as outputs to build an ensemble model to approximate the general controller. A single fixed size least squares support vector machines model and a deep belief networks model are compared with the ensemble modeling method. The experimental results show that the ensemble modeling method can get better results, and it is feasible to construct a general controller for quadrotor certain type of mission.
A dual encoder-based approach to predicting stock price by leveraging online social network
CUI Wenquan, WANG Qingfang
2020, 50(8): 1093-1101. doi: 10.3969/j.issn.0253-2778.2020.08.008
We propose a dual-encoder which encodes the investor sentiment and technical indicators separately to improve the accuracy of the encoder-decoder model in predicting stock price by using two types of information. For the dual-encoder and decoder, we revise the gated recurrent unit (GRU) by removing the reset gate, using the update gate instead of the reset gate function and replacing tanh activation function with ReLU activation function to improve the speed of network training and the accuracy of the model. We regard market sentiment as a discrete-time stochastic process. When fixed time, market sentiment is a variable subject to a certain probability distribution. Sentiment score formulas are built for investor sentiment by a pseudo-label based sentiment classifier, and the market sentiment is estimated through ensemble Bagging learning. The orthogonal table experiment design is used to select parameters in our dual-encoder based model, which greatly reduces the time of parameter adjustment. Finally, experiments are conducted to show that our dual-encoder based model is more accurate than encoder-decoder model, and investor sentiment helps improve the stock forecasting in our model.
Adaptive functional connectivity network learning and application in brain disorders identification
SUN Lei, ZHANG Yining, XUE Yanfang, QIAO Lishan, ZHANG Limei
2020, 50(8): 1102-1109. doi: 10.3969/j.issn.0253-2778.2020.08.009
In recent years, functional connectivity networks (FCN) based on functional magnetic resonance imaging (fMRI) have provided an important tool for the early intervention of brain disorders, such as Alzheimer's disorder (AD) and Autism spectrum disorder (ASD). However, the obtained data are inevitably introduced into structural noises due to participants’ breath, heartbeats and head motions during the scan, which often brings great challenges to the final construction of FCNs. Although conventional data preprocessing methods have been utilized to improve the quality of the data, they still operate in the original data space and separate the data denoising from the FCNs estimation, and thus breaking the internal connection between two steps. Researches shows that data in a certain transform domain may be low-noisy and more informative. Inspired by the transform domain, we propose an adaptive brain network learning model in the light of the transform domain (TD-FCN), which not only improves the quality of the observed data, but also learns the adaptive brain graph in a single framework simultaneously. To verify the effectiveness of the proposed method, we conduct experiments on two public datasets (i.e., ADNI and ABIDE) to identify the patients with mild cognitive impairments (MCIs) and ASDs from health controls (HCs). Experimental results demonstrate that the proposed approach yields statistically significant improvement in multiple performance metrics over traditional methods.
Research on outlier detection algorithm of XmR control chart
CHEN Lifang, WANG Rongjie, LIU Yunqing, ZHOU Xu
2020, 50(8): 1110-1115. doi: 10.3969/j.issn.0253-2778.2020.08.010
A novel outlier detection algorithm was proposed based on the XmR control chart to address the complicated calculation and its time-consuming method in detecting isolated forest anomalies. By calculating the single-valued mean, its moving range and average of the sample attributes, we can draw the control limits and centerlines of the X and mR charts, and the single-valued attributes of the samples in the chart. According to the points in the X chart that exceeds the limits Sample number, add 1 to the sample number corresponding to the point that exceeds the limit in the mR graph, we take the union and delete it from the data, and then replace them after the deletion of the anomaly point with the CART. We use the random forest and support vector machine algorithm for experimental validations. The results show that this method has a faster speed and better precisions compared with the isolation forest method, which provides a new research idea for outlier detection.
Preparation and adsorption property of poly(acrylic acid)-based polyHIPEs through polymerization of O/W HIPE induced by γ-ray radiation and REDOX system
SONG Yuanrui, LIU Huarong, CUI Xiaoling, TAI Chen, CHEN Weijian, LYU Zhijun
2020, 50(8): 1116-1123. doi: 10.3969/j.issn.0253-2778.2020.08.011
Porous poly(acrylic acid)-based polyHIPEs were prepared by the polymerization of oil-in-water high internal phase emulsions (O/W HIPEs) initiated by γ-ray radiation and redox system, respectively. Compared with normal dried gel materials used as hydrogels, polyHIPEs prepared from O/W HIPEs have higher porosity and a more uniform pore structure, and their pore sizes can be adjusted by changing internal phase volume. Observed by scanning electron microscope (SEM), polyHIPEs obtained by γ-ray radiation polymerization were found to have a more ordered and completly porous structure than those obtained by redox initiation system. The adsorption tests of methylene blue (MB) on different samples showed that compared with those prepared by chemical method, polyHIPEs obtained by radiation method had a faster adsorption rate and a higher saturation adsorption capacity. The sample R-25-48 of γ-ray radiation polymerization for 48 h with an absorption dose of 151.2 kGy could adsorb up to 1175.9 mg of methylene blue per gram in 250 mg/L methylene blue solution, and the sample R-25-8 of radiation polymerization for 8 h with an absorbed dose of 25.2 kGy could absorb a lot of water, up to nearly 120 times its original weight, which means that theγ-ray radiation polymerization process affects the structure and properties of polyHIPE materials. The adsorption isotherms and kinetics models were used to fit the adsorption experimental results of MB on polyHIPEs obtained by γ-ray radiation polymerization. It is found that adsorption isotherm fits the Langmuir model and adsorption kinetics conforms to pseudo-second-order model, which indicates that the adsorption process is dominated by chemisorption with the mechanism of monolayer adsorption.
Forecast and analysis of the epidemics trend of COVID-19 in the United States by TRP-SEAMRD model
ZHU Kehang, CHEN Zeying, CHENG Fengyu, TAO Wanyin, ZHU Shu
2020, 50(8): 1124-1133. doi: 10.3969/j.issn.0253-2778.2020.08.012
The traditional SEIR(susceptible-exposed-infectious-recovered/removed) model is a simplified dynamical predictive model which does not consider the impact of changes in the anti-epidemic policy. We take the US anti-epidemic policy and the incubation period characteristic of COVID-19 into account to propose the TRP-SEAMRD(test-restricted-phased SEAMRD) model for the pandemic in US. The model fits well with the figures of COVID-19 infections, recovery and death in the United States during February ~ August 2020. According to the data generated from the model, some of the characteristics of COVID-19 can be abstracted. Based on the TRP-SEAMRD model, we can analyze the impact of the improper anti-epidemic policy at the early stage of the epidemic.The effect of the subsequent “stay at home”epidemic controlling measures is also considered and analyzed. Finally, future development of the pandemic in the US under different degrees of social control is simulated,offering a reference for formulating scientific anti-epidemic measures.
Plasmonic geometric metasurfaces for high-purity polarization conversion
2020, 50(8): 1134-1137. doi: 10.3969/j.issn.0253-2778.2020.08.013
Plasmonic metasurfaces are made up of metallic artificial micro-structures with two-dimensional subwavelength periods, which can realize full control of light via tailoring the wavefronts. Currently, the purity of cross-polarization for transmissive plasmonic metasurfaces is low, leaving that both the signal (cross-polarization) and background (co-polarization) light exist in the transmitted light. Here, a rectangle-hole-based plasmonic metasurface made in a gold film was proposed to realize high-purity conversion of circular polarization. By using the finite-difference time-domain (FDTD) method, the dimension of the rectangle hole was optimized numerically to obtain the theoretical polarization purity of 99.5% in the transmitted light meanwhile maintain the total conversion efficiency larger than 10%. In addition, such a structure has good tolerance to the thickness of film, which benefit its practical applications such as holograms, lenses and gratings.
Study on ultra-low-field nuclear magnetic resonance spectroscopy based on high-sensitivity atomic magnetometer
XU Wenjie, JIANG Min, PENG Xinhua
2020, 50(8): 1138-1143. doi: 10.3969/j.issn.0253-2778.2020.08.014
The frequency of the spectral line and the splitting rule under the ultra-low magnetic field were theoretically given. Then, using the home-built ultra-low-field NMR spectrometer based on a high-sensitivity atomic magnetometer, an experimental study on ultra-low-field NMR spectroscopy was carried out. Taking a typical AXn-type organic molecule as an example, the J-coupling spectra under zero magnetic field was measured and the J-coupling parameters were accurately obtained by combining a theoretical analysis in a variety of organic molecules. For chemical samples with the same spectral structure under zero magnetic field, by applying a weak static magnetic field (nT) to the samples, it was observed that different samples have a unique ultra-low field NMR spectral splitting, which can be used as the "fingerprinting" of the sample to identify them.
Pro-cyclicality of margin trading and short selling based on EMD method
PAN Wanbin, ZHANG Lian
2020, 50(8): 1144-1155. doi: 10.3969/j.issn.0253-2778.2020.08.015
The empirical mode decomposition (EMD) method was applied to measure the cyclical fluctuations of the stock index, and then used to study the pro-cyclicality of China’s margin trading and short selling. The research results show that margin trading has a positive pro-cyclical effect on the overall market, and short selling has a counter-cyclical affect. The vector autoregressive (VAR) model was applied to test the dynamic between margin trading and short selling and stock index. It was found that the impulse response of stock index to margin trading is positive, and the pro-cyclicality of margin trading aggravates the rise and fall of the market; the impulse response of stock index to short selling is negative but not significant, because the scale of securities lending is too small and its counter-cyclical effect can not stabilize the market.
Group stochastic gradient descent: A tradeoff between straggler and staleness
GAO Xiang, CHEN Li
2020, 50(8): 1156-1161. doi: 10.3969/j.issn.0253-2778.2020.08.016
Distributed stochastic gradient descent(DSGD)is widely used for large scale distributed machine learning. Two typical implementations of DSGD are synchronous SGD(SSGD)and asynchronous SGD(ASGD). In SSGD, all workers should wait for each other and the training speed will be slowed down to that of the straggler. In ASGD, the stale gradients can result in a poorly trained model. To solve this problem, a new version of distributed SGD method based named group SGD(GSGD)is proposed, which puts workers with similar computation and communication performance in a group and divides them into several groups. The workers in the same group work in a synchronous manner while different groups work in an asynchronous manner. The proposed method can migrate the straggler problem since workers in the same group spend little time waiting for each other. The staleness of the method is small since the number of groups is much smaller than the number of workers. The convergence of the method is proved through theoretical analysis. Simulation results show that the method converges faster than SSGD and ASGD in the heterogeneous cluster.
UAV target tracking based on visual attention mechanism
LI Peng, ZHENG Yu, ZHANG Tangui
2020, 50(8): 1162-1169. doi: 10.3969/j.issn.0253-2778.2020.08.017
In recent years, the demand for small Unmanned Aerial Vehicles (UAV) in GPS-denied environment is increasingly strong. To solve the problem of multi-target recognition, we study the multi moving target recognition and location technology based on the platform of the small multi-rotor UAV. We used a method to quickly locate the region of interest based on the visual attention mechanism, and then used the machine learning algorithm to classify the region of interest to obtain the target accurately. Our method can track the specified target in the image and locate the target in real time, which the algorithm delay is about 50ms and the location error is less than 15 cm. Our solution can effectively reduce the influence of light variation, motion blur, the color analogue interference and complex background. The ground robot is used as the tracking target to test and verify the algorithm, which can achieve a better tracking effect.
MAEA-DeepLab: A semantic segmentation network with multi-feature attention effective aggregation module
ZHAO Liu, LU Jun, LIU Yang
2020, 50(8): 1170-1180. doi: 10.3969/j.issn.0253-2778.2020.08.018
To realize the low cost of network training, the computational complexity is greatly reduced while maintaining high precision. A semantic segmentation network with multi-feature attention effective aggregation module(MAEA) is proposed: MAEA-DeepLab. A 16 stride low-resolution feature map for down-sampling is adopted in the encoder’s network backbone, and high-level features are obtained. The decoder makes full use of the feature's spatial attention mechanism through the MAEA module, effectively aggregates multiple features, and obtains high-resolution features with strong semantic representation. Then the ability of the decoder to recover important details is effectively improved, and high-precision segmentation is achieved. Multiply-adds in MAEA-DeepLab is 943.02B, only 30.9% of the DeepLabV3+ architecture, which greatly reduces the computational complexity. The architecture is not pre-training on the COCO dataset. It performs semantic semantic segmentation Benchmark tests on the test set of with PASCAL VOC 2012 dataset and CityScapes dataset with only two RTX 2080ti GPUs, and the mlOU scores reach 87.5% and 79.9%, respectively. The experimental results show that good semantic segmentation accuracy is achieved with low computational cost in MAEA-DeepLab.
A simulation of the synthetic aperture radar image based on improved CycleGAN
BAI Jiangbo, YANG Yang, ZHANG Wensheng
2020, 50(8): 1181-1186. doi: 10.3969/j.issn.0253-2778.2020.08.019
The cross-modal data of targets is of great significance to the improvement of the performance of cross-modal detection and multi-modal fusion algorithms based on deep neural networks. Due to the particularity of SAR images, the cost of obtaining paired data is very high, and most of the existing SAR image generation algorithms focus on improving image diversity and small-scale scene generation, and rarely involve image pairing conversion for specific scenes. In this paper, the improved cycle consistency against network CycleGAN is used to achieve the simulation of SAR images of SAR image targets and scenes, and the least square loss is used to improve the network, which improves the network performance and improves the imaging quality. The simulation experiment of SAR image is carried out. The results show that the method produced in this paper has the best fineness and stability, and achieves better simulation results.
A one-shot learning algorithm using support set information during training
XIN Shouyu, ZHENG Ruirui, ZHOU Yu, LIU Wenpeng, HE Jianjun
2020, 50(8): 1187-1192. doi: 10.3969/j.issn.0253-2778.2020.08.020
The purpose of one-shot learning is to use a source category dataset containing a large number of training samples and a target category dataset containing only one training sample per category to construct a learning algorithm that enables accurate classification of samples in the target category space. The existing one-shot learning algorithm mainly uses the source category data to train the model, and then uses the training data of the target category as the support set to realize the classification of the unlabeled samples during the test. Therefore, it fails to effectively utilize the information of the support set during the training. Here, a one-shot learning algorithm using support set information in both the training and test stages is established. The basic idea is to use Siamese neural networks to build models and add support set information during training, that is, to make the similarity between different types of support set samples as small as possible. Experimental results on Omniglot data set and Manchu recognition show that the proposed algorithm can achieve better recognition accuracy.