Most Read articles

    Published in last 1 year |  In last 2 years |  In last 3 years |  All

    All
    Please wait a minute...
    For Selected: Toggle Thumbnails
    Knowledge graph survey: representation, construction, reasoning and knowledge hypergraph theory
    TIAN Ling, ZHANG Jinchuan, ZHANG Jinhao, ZHOU Wangtao, ZHOU Xue
    Journal of Computer Applications    2021, 41 (8): 2161-2186.   DOI: 10.11772/j.issn.1001-9081.2021040662
    Abstract729)      PDF (2811KB)(2031)       Save
    Knowledge Graph (KG) strongly support the research of knowledge-driven artificial intelligence. Aiming at this fact, the existing technologies of knowledge graph and knowledge hypergraph were analyzed and summarized. At first, from the definition and development history of knowledge graph, the classification and architecture of knowledge graph were introduced. Second, the existing knowledge representation and storage methods were explained. Then, based on the construction process of knowledge graph, several knowledge graph construction techniques were analyzed. Specifically, aiming at the knowledge reasoning, an important part of knowledge graph, three typical knowledge reasoning approaches were analyzed, which are logic rule-based, embedding representation-based, and neural network-based. Furthermore, the research progress of knowledge hypergraph was introduced along with heterogeneous hypergraph. To effectively present and extract hyper-relational characteristics and realize the modeling of hyper-relation data as well as the fast knowledge reasoning, a three-layer architecture of knowledge hypergraph was proposed. Finally, the typical application scenarios of knowledge graph and knowledge hypergraph were summed up, and the future researches were prospected.
    Reference | Related Articles | Metrics
    Survey of communication overhead of federated learning
    Xinyuan QIU, Zecong YE, Xiaolong CUI, Zhiqiang GAO
    Journal of Computer Applications    2022, 42 (2): 333-342.   DOI: 10.11772/j.issn.1001-9081.2021020232
    Abstract656)   HTML133)    PDF (1356KB)(1440)       Save

    To solve the irreconcilable contradiction between data sharing demands and requirements of privacy protection, federated learning was proposed. As a distributed machine learning, federated learning has a large number of model parameters needed to be exchanged between the participants and the central server, resulting in higher communication overhead. At the same time, federated learning is increasingly deployed on mobile devices with limited communication bandwidth and limited power, and the limited network bandwidth and the sharply raising client amount will make the communication bottleneck worse. For the communication bottleneck problem of federated learning, the basic workflow of federated learning was analyzed at first, and then from the perspective of methodology, three mainstream types of methods based on frequency reduction of model updating, model compression and client selection respectively as well as special methods such as model partition were introduced, and a deep comparative analysis of specific optimization schemes was carried out. Finally, the development trends of federated learning communication overhead technology research were summarized and prospected.

    Table and Figures | Reference | Related Articles | Metrics
    Path planning method of unmanned aerial vehicle based on chaos sparrow search algorithm
    TANG Andi, HAN Tong, XU Dengwu, XIE Lei
    Journal of Computer Applications    2021, 41 (7): 2128-2136.   DOI: 10.11772/j.issn.1001-9081.2020091513
    Abstract503)      PDF (1479KB)(1053)       Save
    Focusing on the issues of large alculation amount and difficult convergence of Unmanned Aerial Vehicle (UAV) path planning, a path planning method based on Chaos Sparrow Search Algorithm (CSSA) was proposed. Firstly, a two-dimensional task space model and a path cost model were established, and the path planning problem was transformed into a multi-dimensional function optimization problem. Secondly, the cubic mapping was used to initialize the population, and the Opposition-Based Learning (OBL) strategy was used to introduce elite particles, so as to enhance the diversity of the population and expand the scope of the search area. Then, the Sine Cosine Algorithm (SCA) was introduced, and the linearly decreasing strategy was adopted to balance the exploitation and exploration abilities of the algorithm. When the algorithm fell into stagnation, the Gaussian walk strategy was adopted to make the algorithm jump out of the local optimum. Finally, the performance of the proposed improved algorithm was verified on 15 benchmark test functions and applied to solve the path planning problem. Simulation results show that CSSA has better optimization performance than Particle Swarm Optimization (PSO) algorithm, Beetle Swarm Optimization (BSO) algorithm, Whale Optimization Algorithm (WOA), Grey Wolf Optimizer (GWO) algorithm and Sparrow Search Algorithm (SSA), and can quickly obtain a safe and feasible path with optimal cost and satisfying constraints, which proves the effectiveness of the proposed method.
    Reference | Related Articles | Metrics
    Optimized CKKS scheme based on learning with errors problem
    ZHENG Shangwen, LIU Yao, ZHOU Tanping, YANG Xiaoyuan
    Journal of Computer Applications    2021, 41 (6): 1723-1728.   DOI: 10.11772/j.issn.1001-9081.2020091447
    Abstract405)      PDF (760KB)(613)       Save
    Focused on the issue that the CKKS (Cheon-Kim-Kim-Song) homomorphic encryption scheme based on the Learning With Errors (LWE) problem has large ciphertext, complicated calculation key generation and low homomorphic calculation efficiency in the encrypted data calculation, an optimized scheme of LWE type CKKS was proposed through the method of bit discarding and homomorphic calculation key reorganization. Firstly, the size of the ciphertext in the homomorphic multiplication process was reduced by discarding part of the low-order bits of the ciphertext vector and part of the low-order bits of the ciphertext tensor product in the homomorphic multiplication. Secondly, the method of bit discarding was used to reorganize and optimize the homomorphic calculation key, so as to remove the irrelevant extension items in powersof2 during the key exchange procedure and reduce the scale of the calculation key as well as the noise increase in the process of homomorphic multiplication. On the basis of ensuring the security of the original scheme, the proposed optimized scheme makes the dimension of the calculation key reduced, and the computational complexity of the homomorphic multiplication reduced. The analysis results show that the proposed optimized scheme reduces the computational complexity of the homomorphic calculation and calculation key generation process to a certain extent, so as to reduce the storage overhead and improve the efficiency of the homomorphic multiplication operation.
    Reference | Related Articles | Metrics
    Review of deep learning-based medical image segmentation
    CAO Yuhong, XU Hai, LIU Sun'ao, WANG Zixiao, LI Hongliang
    Journal of Computer Applications    2021, 41 (8): 2273-2287.   DOI: 10.11772/j.issn.1001-9081.2020101638
    Abstract385)      PDF (2539KB)(773)       Save
    As a fundamental and key task in computer-aided diagnosis, medical image segmentation aims to accurately recognize the target regions such as organs, tissues and lesions at pixel level. Different from natural images, medical images show high complexity in texture and have the boundaries difficult to judge caused by ambiguity, which is the fault of much noise due to the limitations of the imaging technology and equipment. Furthermore, annotating medical images highly depends on expertise and experience of the experts, thereby leading to limited available annotations in the training and potential annotation errors. For medical images suffer from ambiguous boundary, limited annotated data and large errors in the annotations, which makes it is a great challenge for the auxiliary diagnosis systems based on traditional image segmentation algorithms to meet the demands of clinical applications. Recently, with the wide application of Convolutional Neural Network (CNN) in computer vision and natural language processing, deep learning-based medical segmentation algorithms have achieved tremendous success. Firstly the latest research progresses of deep learning-based medical image segmentation were summarized, including the basic architecture, loss function, and optimization method of the medical image segmentation algorithms. Then, for the limitation of medical image annotated data, the mainstream semi-supervised researches on medical image segmentation were summed up and analyzed. Besides, the studies related to measuring uncertainty of the annotation errors were introduced. Finally, the characteristics summary and analysis as well as the potential future trends of medical image segmentation were listed.
    Reference | Related Articles | Metrics
    Federated learning survey:concept, technology, application and challenge
    Journal of Computer Applications    DOI: 10.11772/j.issn.1001-9081.2021101821
    Accepted: 23 December 2021

    Bamboo strip surface defect detection method based on improved CenterNet
    GAO Qinquan, HUANG Bingcheng, LIU Wenzhe, TONG Tong
    Journal of Computer Applications    2021, 41 (7): 1933-1938.   DOI: 10.11772/j.issn.1001-9081.2020081167
    Abstract373)      PDF (1734KB)(347)       Save
    In bamboo strip surface defect detection, the bamboo strip defects have different shapes and messy imaging environment, and the existing target detection model based on Convolutional Neural Network (CNN) does not take advantage of the neural network when facing such specific data; moreover, the sources of bamboo strips are complicated and there exist other limited conditions, so that it is impossible to collect all types of data, resulting in a small amount of bamboo strip defect data that CNN cannot fully learn. To address these problems, a special detection network aiming at bamboo strip defects was proposed. The basic framework of the proposed network is CenterNet. In order to improve the detection performance of CenterNet in less bamboo strip defect data, an auxiliary detection module based on training from scratch was designed:when the network started training, the CenterNet part that uses the pre-training model was frozen, and the auxiliary detection module was trained from scratch according to the defect characteristics of the bamboo strips; when the loss of the auxiliary detection module stabilized, the module was intergrated with the pre-trained main part by a connection method of attention mechanism. The proposed detection network was trained and tested on the same training sets with CenterNet and YOLO v3 which is currently commonly used in industrial detection. Experimental results show that on the bamboo strip defect detection dataset, the mean Average Precision (mAP) of the proposed method is 16.45 and 9.96 percentage points higher than those of YOLO v3 and CenterNet, respectively. The proposed method can effectively detect the different shaped defects of bamboo strips without increasing too much time consumption, and has a good effect in actual industrial applications.
    Reference | Related Articles | Metrics
    Review of spatio-temporal trajectory sequence pattern mining methods
    KANG Jun, HUANG Shan, DUAN Zongtao, LI Yixiu
    Journal of Computer Applications    2021, 41 (8): 2379-2385.   DOI: 10.11772/j.issn.1001-9081.2020101571
    Abstract372)      PDF (1204KB)(1087)       Save
    With the rapid development of global positioning technology and mobile communication technology, huge amounts of trajectory data appear. These data are true reflections of the moving patterns and behavior characteristics of moving objects in the spatio-temporal environment, and they contain a wealth of information which carries important application values for the fields such as urban planning, traffic management, service recommendation, and location prediction. And the applications of spatio-temporal trajectory data in these fields usually need to be achieved by sequence pattern mining of spatio-temporal trajectory data. Spatio-temporal trajectory sequence pattern mining aims to find frequently occurring sequence patterns from the spatio-temporal trajectory dataset, such as location patterns (frequent trajectories, hot spots), activity periodic patterns, and semantic behavior patterns, so as to mine hidden information in the spatio-temporal data. The research progress of spatial-temporal trajectory sequence pattern mining in recent years was summarized. Firstly, the data characteristics and applications of spatial-temporal trajectory sequence were introduced. Then, the mining process of spatial-temporal trajectory patterns was described:the research situation in this field was introduced from the perspectives of mining location patterns, periodic patterns and semantic patterns based on spatial-temporal trajectory sequence. Finally, the problems existing in the current spatio-temporal trajectory sequence pattern mining methods were elaborated, and the future development trends of spatio-temporal trajectory sequence pattern mining method were prospected.
    Reference | Related Articles | Metrics
    Improved high-dimensional many-objective evolutionary algorithm based on decomposition
    Gangzhu QIAO, Rui WANG, Chaoli SUN
    Journal of Computer Applications    2021, 41 (11): 3097-3103.   DOI: 10.11772/j.issn.1001-9081.2020121895
    Abstract370)   HTML88)    PDF (525KB)(368)       Save

    In the reference vector based high-dimensional many-objective evolutionary algorithms, the random selection of parent individuals will slow down the speed of convergence, and the lack of individuals assigned to some reference vectors will weaken the diversity of population. In order to solve these problems, an Improved high-dimensional Many-Objective Evolutionary Algorithm based on Decomposition (IMaOEA/D) was proposed. Firstly, when a reference vector was assigned at least two individuals in the framework of decomposition strategy, the parent individuals were selected for reproduction of offspring according to the distance from the individual assigned to the reference vector to the ideal point, so as to increase the search speed. Then, for the reference vector that was not assigned at least two individuals, the point with the smallest distance from the ideal point along the reference vector was selected from all the individuals, so that at least two individuals and the reference vector were associated. Meanwhile, by guaranteeing one individual was related to each reference vector after environmental selection, the diversity of population was ensured. The proposed method was tested and compared with other four high-dimensional many-objective optimization algorithms based on decomposition on the MaF test problem sets with 10 and 15 objectives. Experimental results show that, the proposed algorithm has good optimization ability for high-dimensional many-objective optimization problems: the optimization results of the proposed algorithm on 14 test problems of the 30 test problems are better than those of the other four comparison algorithms. Especially, the proposed algorithm has certain advantage on the degradation problem optimization.

    Table and Figures | Reference | Related Articles | Metrics
    Unsupervised attributed graph embedding model based on node similarity
    Yang LI, Anbiao WU, Ye YUAN, Linlin ZHAO, Guoren WANG
    Journal of Computer Applications    2022, 42 (1): 1-8.   DOI: 10.11772/j.issn.1001-9081.2021071221
    Abstract363)   HTML100)    PDF (864KB)(395)       Save

    Attributed graph embedding aims to represent the nodes in an attributed graph into low-dimensional vectors while preserving the topology information and attribute information of the nodes. There are lots of works related to attributed graph embedding. However, most of algorithms proposed in them are supervised or semi-supervised. In practical applications, the number of nodes that need to be labeled is large, which makes these algorithms difficult and consume huge manpower and material resources. Above problems were reanalyzed from an unsupervised perspective, and an unsupervised attributed graph embedding algorithm was proposed. Firstly, the topology information and attribute information of the nodes were calculated respectively by using the existing non-attributed graph embedding algorithm and attributes of the attributed graph. Then, the embedding vector of the nodes was obtained by using Graph Convolutional Network (GCN), and the difference between the embedding vector and the topology information and the difference between the embedding vector and attribute information were minimized. Finally, similar embeddings was obtained by the paired nodes with similar topological information and attribute information. Compared with Graph Auto-Encoder (GAE) method, the proposed method has the node classification accuracy improved by 1.2 percentage points and 2.4 percentage points on Cora and Citeseer datasets respectively. Experimental results show that the proposed method can effectively improve the quality of the generated embedding.

    Table and Figures | Reference | Related Articles | Metrics
    Sequential multimodal sentiment analysis model based on multi-task learning
    ZHANG Sun, YIN Chunyong
    Journal of Computer Applications    2021, 41 (6): 1631-1639.   DOI: 10.11772/j.issn.1001-9081.2020091416
    Abstract333)      PDF (1150KB)(801)       Save
    Considering the issues of unimodal feature representation and cross-modal feature fusion in sequential multimodal sentiment analysis, a multi-task learning based sentiment analysis model was proposed by combining with multi-head attention mechanism. Firstly, Convolution Neural Network (CNN), Bidirectional Gated Recurrent Unit (BiGRU) and Multi-Head Self-Attention (MHSA) were used to realize the sequential unimodal feature representation. Secondly, the bidirectional cross-modal information was fused by multi-head attention. Finally, based on multi-task learning, the sentiment polarity classification and sentiment intensity regression were added as auxiliary tasks to improve the comprehensive performance of the main task of sentiment score regression. Experimental results demonstrate that the proposed model improves the accuracy of binary classification by 7.8 percentage points and 3.1 percentage points respectively on CMU Multimodal Opinion Sentiment and Emotion Intensity (CMU-MOSEI) and CMU Multimodal Opinion level Sentiment Intensity (CMU-MOSI) datasets compared with multimodal factorization model. Therefore, the proposed model is applicable for the sentiment analysis problems under multimodal scenarios, and can provide the decision supports for product recommendation, stock market forecasting, public opinion monitoring and other relevant applications.
    Reference | Related Articles | Metrics
    Survey of anonymity and tracking technology in Monero
    Dingkang LIN, Jiaqi YAN, Nandeng BA, Zhenhao FU, Haochen JIANG
    Journal of Computer Applications    2022, 42 (1): 148-156.   DOI: 10.11772/j.issn.1001-9081.2021020296
    Abstract327)   HTML21)    PDF (723KB)(300)       Save

    Virtual digital currency provides a breeding ground for terrorist financing, money laundering, drug trafficking and other criminal activities. As a representative emerging digital currency, Monero has a universally acknowledged high anonymity. Aiming at the problem of using Monroe anonymity to commit crimes, Monero anonymity technology and tracking technology were explored as well as the research progresses were reviewed in recent years, so as to provide technical supports for effectively tackling the crimes based on blockchain technology. In specific, the evolution of Monero anonymity technology was summarized, and the tracking strategies of Monero anonymity technology in academic circles were sorted out. Firstly, in the anonymity technologies, ring signature, guaranteed unlinkability (one-off public key), guaranteed untraceability, and the important version upgrading for improving anonymity were introduced. Then, in tracking technologies, the attacks such as zero mixin attack, output merging attack, guess-newest attack, closed set attack, transaction flooding attack, tracing attacks from remote nodes and Monero ring attack were introduced. Finally, based on the analysis of anonymity technologies and tracking strategies, four conclusions were obtained: the development of anonymity technology and the development of tracking technology of Monero promote each other; the application of Ring Confidential Transactions (RingCT) is a two-edged sword, which makes the passive attack methods based on currency value ineffective, and also makes the active attack methods easier to succeed; output merging attack and zero mixin attack complement each other; Monero’s system security chain still needs to be sorted out.

    Table and Figures | Reference | Related Articles | Metrics
    Research advances in disentangled representation learning
    Keyang CHENG, Chunyun MENG, Wenshan WANG, Wenxi SHI, Yongzhao ZHAN
    Journal of Computer Applications    2021, 41 (12): 3409-3418.   DOI: 10.11772/j.issn.1001-9081.2021060895
    Abstract308)   HTML41)    PDF (877KB)(207)       Save

    The purpose of disentangled representation learning is to model the key factors that affect the form of data, so that the change of a key factor only causes the change of data on a certain feature, while the other features are not affected. It is conducive to face the challenge of machine learning in model interpretability, object generation and operation, zero-shot learning and other issues. Therefore, disentangled representation learning always be a research hotspot in the field of machine learning. Starting from the history and motives of disentangled representation learning, the research status and applications of disentangled representation learning were summarized, the invariance, reusability and other characteristics of disentangled representation learning were analyzed, and the research on the factors of variation via generative entangling, the research on the factors of variation with manifold interaction, and the research on the factors of variation using adversarial training were introduced, as well as the latest research trends such as a Variational Auto-Encoder (VAE) named β-VAE were introduced. At the same time, the typical applications of disentangled representation learning were shown, and the future research directions were prospected.

    Table and Figures | Reference | Related Articles | Metrics
    Review of remote sensing image change detection
    REN Qiuru, YANG Wenzhong, WANG Chuanjian, WEI Wenyu, QIAN Yunyun
    Journal of Computer Applications    2021, 41 (8): 2294-2305.   DOI: 10.11772/j.issn.1001-9081.2020101632
    Abstract279)      PDF (1683KB)(582)       Save
    As a key technology of land use/land cover detection, change detection aims to detect the changed part and its type in the remote sensing data of the same region in different periods. In view of the problems in traditional change detection methods, such as heavy manual labor and poor detection results, a large number of change detection methods based on remote sensing images have been proposed. In order to further understand the change detection technology based on remote sensing images and further study on the change detection methods, a comprehensive review of change detection was carried out by sorting, analyzing and comparing a large number of researches on change detection. Firstly, the development process of change detection was described. Then, the research progress of change detection was summarized in detail from three aspects:data selection and preprocessing, change detection technology, post-processing and precision evaluation, where the change detection technology was mainly summarized from analysis unit and comparison method respectively. Finally, the summary of the problems in each stage of change detection was performed and the future development directions were proposed.
    Reference | Related Articles | Metrics
    Dual-channel night vision image restoration method based on deep learning
    NIU Kangli, CHEN Yuzhang, SHEN Junfeng, ZENG Zhangfan, PAN Yongcai, WANG Yichong
    Journal of Computer Applications    2021, 41 (6): 1775-1784.   DOI: 10.11772/j.issn.1001-9081.2020091411
    Abstract268)      PDF (1916KB)(486)       Save
    Due to the low light level and low visibility of night scene, there are many problems in night vision image, such as low signal to noise ratio and low imaging quality. To solve the problems, a dual-channel night vision image restoration method based on deep learning was proposed. Firstly, two Convolutional Neural Network (CNN) based on Fully connected Multi-scale Residual learning Block (FMRB) were used to extract multi-scale features and fuse hierarchical features of infrared night vision images and low-light-level night vision images respectively, so as to obtain the reconstructed infrared image and enhanced low-light-level image. Then, the two processed images were fused by the adaptive weighted averaging algorithm, and the effective information of the more salient one in the two images was highlighted adaptively according to the different scenes. Finally, the night vision restoration images with high resolution and good visual effect were obtained. The reconstructed infrared night vision image obtained by the FMRB based deep learning network had the average values of Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM) by 3.56 dB and 0.091 2 higher than the image obtained by Super-Resolution Convolutional Neural Network (SRCNN) reconstruction algorithm respectively, and the enhanced low-light-level night vision image obtained by the FMRB based deep learning network had the average values of PSNR and SSIM by 6.82dB and 0.132 1 higher than the image obtained by Multi-Scale Retinex with Color Restoration (MSRCR). Experimental results show that, by using the proposed method, the resolution of reconstructed image is improved obviously and the brightness of the enhanced image is also improved significantly, and the visual effect of the fusion image obtained by the above two images is better. It can be seen that the proposed algorithm can effectively restore the night vision images.
    Reference | Related Articles | Metrics
    Research progress on binary code similarity search
    Bing XIA, Jianmin PANG, Xin ZHOU, Zheng SHAN
    Journal of Computer Applications    2022, 42 (4): 985-998.   DOI: 10.11772/j.issn.1001-9081.2021071267
    Abstract259)   HTML47)    PDF (841KB)(288)       Save

    With the rapid development of Internet of Things (IoT) and industrial Internet, the research of cyberspace security has been paid more and more attention by industry and academia. Because the source code cannot be obtained, binary code similarity search has become a key core technology for vulnerability mining and malware code analysis. Firstly, the basic concepts of binary code similarity search and the framework of binary code similarity search system were introduced. Secondly, the development status of binary code technology about syntax similarity search, semantic similarity search and pragmatic similarity search were discussed. Then, the existing solutions were summarized and compared from the perspectives of binary hash, instruction sequence, graph structure, basic block semantics, feature learning, debugging information recovery and advanced semantic recognition of functions. Finally, the future development direction and prospect of binary code similarity search were looked forward to.

    Table and Figures | Reference | Related Articles | Metrics
    Text multi-label classification method incorporating BERT and label semantic attention
    Xueqiang LYU, Chen PENG, Le ZHANG, Zhi’an DONG, Xindong YOU
    Journal of Computer Applications    2022, 42 (1): 57-63.   DOI: 10.11772/j.issn.1001-9081.2021020366
    Abstract255)   HTML22)    PDF (577KB)(485)       Save

    Multi-Label Text Classification (MLTC) is one of the important subtasks in the field of Natural Language Processing (NLP). In order to solve the problem of complex correlation between multiple labels, an MLTC method TLA-BERT was proposed by incorporating Bidirectional Encoder Representations from Transformers (BERT) and label semantic attention. Firstly, the contextual vector representation of the input text was learned by fine-tuning the self-coding pre-training model. Secondly, the labels were encoded individually by using Long Short-Term Memory (LSTM) neural network. Finally, the contribution of text to each label was explicitly highlighted with the use of an attention mechanism in order to predict the multi-label sequences. Experimental results show that compared with Sequence Generation Model (SGM) algorithm, the proposed method improves the F value by 2.8 percentage points and 1.5 percentage points on the Arxiv Academic Paper Dataset (AAPD) and Reuters Corpus Volume I (RCV1)-v2 public dataset respectively.

    Table and Figures | Reference | Related Articles | Metrics
    Annotation method for joint extraction of domain-oriented entities and relations
    WU Saisai, LIANG Xiaohe, XIE Nengfu, ZHOU Ailian, HAO Xinning
    Journal of Computer Applications    2021, 41 (10): 2858-2863.   DOI: 10.11772/j.issn.1001-9081.2020101678
    Abstract251)      PDF (803KB)(373)       Save
    In view of the problems of low efficiency, error propagation, and entity redundancy in traditional entities and relations annotation methods, and for the fact that there is the characteristic of "the overlapping relationship between one entity (main-entity) and multiple entities at the same time" in corpuses of some domains, a new annotation method for joint extraction of domain entities and relations was proposed. First, the main entity was marked as a fixed label, each other entity in the text that has relation with the main-entity was marked as the type of relation between the corresponding two entities. This way that entities and relations were simultaneously labeled was able to save at least half of the cost of annotation. Then, the triples were modeled directly instead of modeling entities and relations separately, and, the triple data were able to be obtained through label matching and mapping, which alleviated the problems of overlapping relation extraction, entity redundancy, and error propagation. Finally, the field of crop diseases and pests was taken as the example to conduct experiments, and the Bidirectional Encoder Representations from Transformers (BERT)-Bidirectional Long Short-Term Memory (BiLSTM)+Conditional Random Field (CRF) end-to-end model was tested the performance on the dataset of 1 619 crop diseases and pests articles. Experimental results show that this model has the F1 value 47.83 percentage points higher than the pipeline method based on the traditional annotation method+BERT model; compared with the joint learning method based on the new annotation method+BiLSTM+CRF model, Convolutional Neural Network (CNN)+BiLSTM+CRF or other classic models, the F1 value of the model increased by 9.55 percentage points and 10.22 percentage points respectively, which verify the effectiveness of the proposed annotation method and model.
    Reference | Related Articles | Metrics
    Difference detection method of adversarial samples oriented to deep learning
    WANG Shuyan, HOU Zeyu, SUN Jiaze
    Journal of Computer Applications    2021, 41 (7): 1849-1856.   DOI: 10.11772/j.issn.1001-9081.2020081282
    Abstract247)      PDF (2685KB)(357)       Save
    Deep Neural Network (DNN) is proved to be vulnerable to adversarial sample attacks in many key deep learning systems such as face recognition and intelligent driving. And the detection of various types of adversarial samples has problems of insufficient detection and low detection efficiency. Therefore, a deep learning model oriented adversarial sample difference detection method was proposed. Firstly, the residual neural network model commonly used in industrial production was constructed as the model of the adversarial sample generation and detection system. Then, multiple kinds of adversarial attacks were used to attack the deep learning model to generate adversarial sample groups. Finally, a sample difference detection system was constructed, containing total 7 adversarial sample difference detection methods in sample confidence detection, perception detection and anti-interference degree detection. Empirical research was carried out by the constructed method on the MNIST and Cifar-10 datasets. The results show that the adversarial samples belonging to different adversarial attacks have obvious differences in the performance detection on confidence, perception and anti-interference degrees, for example, in the detection of confidence and anti-interference, the adversarial samples with excellent performance indicators in perception show significant insufficiencies compared to other types of adversarial samples. At the same time, it is proved that there is consistency of the differences in the two datasets. By using this detection method, the comprehensiveness and diversity of the model's detection of adversarial samples can be effectively improved.
    Reference | Related Articles | Metrics
    Authorship identification of text based on attention mechanism
    ZHANG Yang, JIANG Minghu
    Journal of Computer Applications    2021, 41 (7): 1897-1901.   DOI: 10.11772/j.issn.1001-9081.2020101528
    Abstract245)      PDF (795KB)(371)       Save
    The accuracy of authorship identification based on deep neural network decreases significantly when faced with a large number of candidate authors. In order to improve the accuracy of authorship identification, a neural network consisting of fast text classification (fastText) and an attention layer was proposed, and it was combined with the continuous Part-Of-Speech (POS) n-gram features for authorship identification of Chinese novels. Compared with Text Convolutional Neural Network (TextCNN), Text Recurrent Neural Network (TextRNN), Long Short-Term Memory (LSTM) network and fastText, the experimental results show that the proposed model obtains the highest classification accuracy. Compared with the fastText model, the introduction of attention mechanism increases the accuracy corresponding to different POS n-gram features by 2.14 percentage points on average; meanwhile, the model retains the high-speed and efficiency of fastText, and the text features used by it can be applied to other languages.
    Reference | Related Articles | Metrics
    Efficient storage scheme for deadline aware distributed matrix multiplication
    Yongzhu ZHAO, Weidong LI, Bin TANG, Feng MEI, Wenda LU
    Journal of Computer Applications    2020, 40 (2): 311-315.   DOI: 10.11772/j.issn.1001-9081.2019091640
    Abstract240)   HTML11)    PDF (742KB)(439)       Save

    Distributed matrix multiplication is a fundamental operation in many distributed machine learning and scientific computing applications, but its performance is greatly influenced by the stragglers commonly existed in the systems. Recently, researchers have proposed a fountain code based coded matrix multiplication method, which can effectively mitigate the effect of stragglers by fully exploiting the partial results of stragglers. However, it lacks the consideration of the storage cost of worker nodes. By considering the tradeoff relationship between the storage cost and the finish time of computation, the computational deadline-aware storage optimization problem for heterogeneous worker nodes was proposed firstly. Then, through the theoretical analysis, the solution based on expectation approximation was presented, and the problem was transformed into a convex optimization problem by relaxation for efficient solution. Simulation results show that in the case of ensuring a large task success rate, the storage overhead of the proposed scheme will rapidly decrease as the task duration is relaxed, and the scheme can greatly reduce the storage overhead brought by encoding. In other words, the proposed scheme can significantly reduce the extra storage overhead while guaranteeing that the whole computation can be finished before the deadline with high probability.

    Table and Figures | Reference | Related Articles | Metrics
    Safety helmet wearing detection algorithm based on improved YOLOv5
    Jin ZHANG, Peiqi QU, Cheng SUN, Meng LUO
    Journal of Computer Applications    2022, 42 (4): 1292-1300.   DOI: 10.11772/j.issn.1001-9081.2021071246
    Abstract237)   HTML8)    PDF (7633KB)(164)       Save

    Aiming at the problems of strong interference and low detection precision of the existing safety helmet wearing detection, an algorithm of safety helmet detection based on improved YOLOv5 (You Only Look Once version 5) model was proposed. Firstly, for the problem of different sizes of safety helmets, the K-Means++ algorithm was used to redesign the size of the anchor box and match it to the corresponding feature layer. Secondly, the multi-spectral channel attention module was embedded in the feature extraction network to ensure that the network was able to learn the weight of each channel autonomously and enhance the information dissemination between the features, thereby strengthening the network ability to distinguish foreground and background. Finally, images of different sizes were input randomly during the training iteration process to enhance the generalization ability of the algorithm. Experimental results show as follows: on the self-built safety helmet wearing detection dataset, the proposed algorithm has the mean Average Precision (mAP) reached 96.0%, the the Average Precision (AP) of workers wearing safety helmet reached 96.7%, and AP of workers without safety helmet reached 95.2%. Compared with the YOLOv5 algorithm, the proposed algorithm has the mAP of helmet safety-wearing detection increased by 3.4 percentage points, and it meets the accuracy requirement of helmet safety-wearing detection in construction scenarios.

    Table and Figures | Reference | Related Articles | Metrics
    Human skeleton-based action recognition algorithm based on spatiotemporal attention graph convolutional network model
    LI Yangzhi, YUAN Jiazheng, LIU Hongzhe
    Journal of Computer Applications    2021, 41 (7): 1915-1921.   DOI: 10.11772/j.issn.1001-9081.2020091515
    Abstract232)      PDF (1681KB)(527)       Save
    Aiming at the problem that the existing human skeleton-based action recognition algorithms cannot fully explore the temporal and spatial characteristics of motion, a human skeleton-based action recognition algorithm based on Spatiotemporal Attention Graph Convolutional Network (STA-GCN) model was proposed, which consisted of spatial attention mechanism and temporal attention mechanism. The spatial attention mechanism used the instantaneous motion information of the optical flow features to locate the spatial regions with significant motion on the one hand, and introduced the global average pooling and auxiliary classification loss during the training process to enable the model to focus on the non-motion regions with discriminability ability on the other hand. While the temporal attention mechanism automatically extracted the discriminative time-domain segments from the long-term complex video. Both of spatial and temporal attention mechanisms were integrated into a unified Graph Convolution Network (GCN) framework to enable the end-to-end training. Experimental results on Kinetics and NTU RGB+D datasets show that the proposed algorithm based on STA-GCN has strong robustness and stability, and compared with the benchmark algorithm based on Spatial Temporal Graph Convolutional Network (ST-GCN) model, the Top-1 and Top-5 on Kinetics are improved by 5.0 and 4.5 percentage points, respectively, and the Top-1 on CS and CV of NTU RGB+D dataset are also improved by 6.2 and 6.7 percentage points, respectively; it also outperforms the current State-Of-the-Art (SOA) methods in action recognition, such as Res-TCN (Residue Temporal Convolutional Network), STA-LSTM, and AS-GCN (Actional-Structural Graph Convolutional Network). The results indicate that the proposed algorithm can better meet the practical application requirements of human action recognition.
    Reference | Related Articles | Metrics
    Knowledge graph driven recommendation model of graph neural network
    LIU Huan, LI Xiaoge, HU Likun, HU Feixiong, WANG Penghua
    Journal of Computer Applications    2021, 41 (7): 1865-1870.   DOI: 10.11772/j.issn.1001-9081.2020081254
    Abstract232)      PDF (991KB)(486)       Save
    The abundant structure and association information contained in Knowledge Graph (KG) can not only alleviate the data sparseness and cold-start in the recommender systems, but also make personalized recommendation more accurately. Therefore, a knowledge graph driven end-to-end recommendation model of graph neural network, named KGLN, was proposed. First, a signal-layer neural network framework was used to fuse the features of individual nodes in the graph, then the aggregation weights of different neighbor entities were changed by adding influence factors. Second, the single-layer was extended to multi-layer by iteration, so that the entities were able to obtain abundant multi-order associated entity information. Finally, the obtained features of entities and users were integrated to generate the prediction score for recommendation. The effects of different aggregation methods and influence factors on the recommendation results were analyzed. Experimental results show that on the datasets MovieLen-1M and Book-Crossing, compared with the benchmark methods such as Factorization Machine Library (LibFM), Deep Factorization Machine (DeepFM), Wide&Deep and RippleNet, KGLN obtains an AUC (Area Under ROC (Receiver Operating Characteristic) curve) improvement of 0.3%-5.9% and 1.1%-8.2%, respectively.
    Reference | Related Articles | Metrics
    Attention-based object detection with millimeter wave radar-lidar fusion
    LI Chao, LAN Hai, WEI Xian
    Journal of Computer Applications    2021, 41 (7): 2137-2144.   DOI: 10.11772/j.issn.1001-9081.2020081334
    Abstract228)      PDF (1710KB)(252)       Save
    To address problems of missing occluded objects, distant objects and objects in extreme weather scenarios when using lidar for object detection in autonomous driving, an attention-based object detection method with millimeter wave radar-lidar feature fusion was proposed. Firstly, the scan frame data of millimeter wave radar and lidar were aggregated into their respective labeled frames, and the points of millimeter wave radar and lidar were spatially aligned, then PointPillar was employed to encode both the millimeter wave radar and lidar data into pseudo images. Finally, the features of both millimeter wave radar and lidar sensors were extracted by the middle convolution layer, and the features maps of them were fused by attention mechanism, and the fused feature map was passed through a single-stage detector to obtain detection results. Experimental results on nuScenes dataset show that compared to the basic PointPillar network, the mean Average Precision (mAP) of the proposed attention fusion algorithm is higher, which performs better than concatenation fusion, multiply fusion and add fusion methods. The visualization results show that the proposed method is effective and can improve the robustness of the network for detecting occluded objects, distant objects and objects surrounded by rain and fog.
    Reference | Related Articles | Metrics
    Plant leaf disease recognition method based on lightweight convolutional neural network
    JIA Heming, LANG Chunbo, JIANG Zichao
    Journal of Computer Applications    2021, 41 (6): 1812-1819.   DOI: 10.11772/j.issn.1001-9081.2020091471
    Abstract227)      PDF (1486KB)(276)       Save
    Aiming at the problems of low accuracy and poor real-time performance of plant leaf disease recognition in the field of agricultural information, a plant leaf disease recognition method based on lightweight Convolutional Neural Network (CNN) was proposed. The Depthwise Separable Convolution (DSC) and Global Average Pooling (GAP) methods were introduced in the original network to replace the standard convolution operation and the fully connected layer part at the end of the network respectively. At the same time, the technique of batch normalization was also applied to the process of training network to improve the intermediate layer data distribution and increase the convergence speed. In order to comprehensively and reliably evaluate the performance of the proposed method, experiments were conducted on the open plant leaf disease image dataset PlantVillage, and loss function convergence curve, test accuracy, parameter memory demand and other indicators were selected to verify the effectiveness of the improved strategy. Experimental results show that the improved network has higher disease recognition accuracy (99.427%) and smaller memory space occupation (6.47 MB), showing that it is superior to other leaf recognition technologies based on neural network, and has strong engineering practicability.
    Reference | Related Articles | Metrics
    Fall behavior detection algorithm for the elderly based on AlphaPose optimization model
    Jingqi MA, Huan LEI, Minyi CHEN
    Journal of Computer Applications    2022, 42 (1): 294-301.   DOI: 10.11772/j.issn.1001-9081.2021020331
    Abstract224)   HTML13)    PDF (7482KB)(400)       Save

    In order to detect the elderly fall high-risk behaviors quickly and accurately on the low-power and low-cost hardware platform, an abnormal behavior detection algorithm based on AlphaPose optimization model was proposed. Firstly, the pedestrian target detection model and pose estimation model were optimized to accelerate the human target detection and pose joint point reasoning. Then, the image coordinate data of human pose joint points were computed rapidly through the optimized AlphaPose model. Finally, the relationship between the head joint point linear velocity and the crotch joint linear velocity at the moment the human body falls was calculated, as well as the change of the angle between the midperpendicular of the torso and X-axis of the image, were calculated to determine the occurrence of the fall. The proposed algorithm was deployed to the Jetson Nano embedded development board, and compared with several main fall detection algorithms based on human pose at present: YOLO (You Only Look Once)v3+Pose, YOLOv4+Pose, YOLOv5+Pose, trt_pose and NanoDet+Pose. Experimental results show that on the used embedded platform when the image resolution is 320×240, the proposed algorithm has the detection frame rate of 8.83 frame/s and the accuracy of 0.913, which are both better than those of the algorithms compared above. The proposed algorithm has relatively high real-time performance and accuracy, and can timely detect the occurrence of the elderly fall behaviors.

    Table and Figures | Reference | Related Articles | Metrics
    Automated English essay scoring method based on multi-level semantic features
    ZHOU Xianbing, FAN Xiaochao, REN Ge, YANG Yong
    Journal of Computer Applications    2021, 41 (8): 2205-2211.   DOI: 10.11772/j.issn.1001-9081.2020101572
    Abstract214)      PDF (935KB)(230)       Save
    The Automated Essay Scoring (AES) technology can automatically analyze and score the essay, and has become one of the hot research problems in the application of natural language processing technology in the education field. Aiming at the current AES methods that separate deep and shallow semantic features, and ignore the impact of multi-level semantic fusion on essay scoring, a neural network model based on Multi-Level Semantic Features (MLSF) was proposed for AES. Firstly, Convolutional Neural Network (CNN) was used to capture local semantic features, and the hybrid neural network was used to capture global semantic features, so that the essay semantic features were obtained from a deep level. Secondly, the feature of the topic layer was obtained by using the essay topic vector of text level. At the same time, aiming at the grammatical errors and language richness features that are difficult to mine by deep learning model, a small number of artificial features were constructed to obtain the linguistic features of the essay from the shallow level. Finally, the essay was automatically scored through the feature fusion. Experimental results show that the proposed model improves the performance significantly on all subsets of the public dataset of the Kaggle ASAP (Automated Student Assessment Prize) champion, with the average Quadratic Weighted Kappa (QWK) of 79.17%, validating the effectiveness of the model in the AES tasks.
    Reference | Related Articles | Metrics
    Multi-head attention memory network for short text sentiment classification
    Yu DENG, Xiaoyu LI, Jian CUI, Qi LIU
    Journal of Computer Applications    2021, 41 (11): 3132-3138.   DOI: 10.11772/j.issn.1001-9081.2021010040
    Abstract211)   HTML20)    PDF (681KB)(143)       Save

    With the development of social networks, it has important social value to analyze the sentiments of massive texts in the social networks. Different from ordinary text classification, short text sentiment classification needs to mine the implicit sentiment semantic features, so it is very difficult and challenging. In order to obtain short text sentiment semantic features at a higher level, a new Multi-head Attention Memory Network (MAMN) was proposed for sentiment classification of short texts. Firstly, n-gram feature information and Ordered Neurons Long Short-Term Memory (ON-LSTM) network were used to improve the multi-head self-attention mechanism to fully extract the internal relationship of the text context, so that the model was able obtain richer text feature information. Secondly, multi-head attention mechanism was adopted to optimize the multi-hop memory network structure, so as to expand the depth of the model and mine higher level contextual internal semantic relations at the same time. A large number of experiments were carried out on Movie Review dataset (MR), Stanford Sentiment Treebank (SST)-1 and SST-2 datasets. The experimental results show that compared with the baseline models based on Recurrent Neural Network (RNN) and Convolutional Neural Network (CNN) structure and some latest works, the proposed MAMN achieves the better classification results, and the importance of multi-hop structure in performance improvement is verified.

    Table and Figures | Reference | Related Articles | Metrics
    Network embedding method based on multi-granularity community information
    Jun HU, Zhengkang XU, Li LIU, Fujin ZHONG
    Journal of Computer Applications    2022, 42 (3): 663-670.   DOI: 10.11772/j.issn.1001-9081.2021040790
    Abstract211)   HTML50)    PDF (758KB)(217)       Save

    Most of the existing network embedding methods only preserve the local structure information of the network, while they ignore other potential information in the network. In order to preserve the community information of the network and reflect the multi-granularity characteristics of the network community structure, a network Embedding method based on Multi-Granularity Community information (EMGC) was proposed. Firstly, the network’s multi-granularity community structure was obtained, the node embedding and the community embedding were initialized. Then, according to the node embedding at previous level of granularity and the community structure at this level of granularity, the community embedding was updated, and the corresponding node embedding was adjusted. Finally, the node embeddings under different community granularities were spliced to obtain the network embedding that fused the community information of different granularities. Experiments on four real network datasets were carried out. Compared with the methods that do not consider community information (DeepWalk, node2vec) and the methods that consider single-granularity community information (ComE, GEMSEC), EMGC’s AUC value on link prediction and F1 score on node classification are generally better than those of the comparison methods. The experimental results show that EMGC can effectively improve the accuracy of subsequent link prediction and node classification.

    Table and Figures | Reference | Related Articles | Metrics
    Tire defect detection method based on improved Faster R-CNN
    WU Zeju, JIAO Cuijuan, CHEN Liang
    Journal of Computer Applications    2021, 41 (7): 1939-1946.   DOI: 10.11772/j.issn.1001-9081.2020091488
    Abstract208)      PDF (1816KB)(376)       Save
    The defects such as sidewall foreign matter, crown foreign body, air bubble, crown split and sidewall root opening that appear in the process of tire production will affect the use of tires after leaving factory, so it is necessary to carry out nondestructive testing on each tire before leaving the factory. In order to achieve automatic detection of tire defects in industry, an automatic tire defect detection method based on improved Faster Region-Convolutional Neural Network (Faster R-CNN) was proposed. Firstly, at the preprocessing stage, the gray level of tire image was stretched by the histogram equalization method to enhance the contrast of the dataset, resulting in a significant difference between gray values of the image target and the background. Secondly, to improve the accuracy of position detection and identification of tire defects, the Faster R-CNN structure was improved. That is the convolution features of the third layer and the convolution features of the fifth layer in ZF (Zeiler and Fergus) convolutional neural network were combined together and output as the input of the regional proposal network layer. Thirdly, the Online Hard Example Mining (OHEM) algorithm was introduced after the RoI (Region-of-Interesting) pooling layer to further improve the accuracy of defect detection. Experimental results show that the tire X-ray image defects can be classified and located accurately by the improved Faster R-CNN defect detection method with average test recognition of 95.7%. In addition, new detection models can be obtained by fine-tuning the network to detect other types of defects..
    Reference | Related Articles | Metrics
    Motion control method of two-link manipulator based on deep reinforcement learning
    WANG Jianping, WANG Gang, MAO Xiaobin, MA Enqi
    Journal of Computer Applications    2021, 41 (6): 1799-1804.   DOI: 10.11772/j.issn.1001-9081.2020091410
    Abstract208)      PDF (875KB)(420)       Save
    Aiming at the motion control problem of two-link manipulator, a new control method based on deep reinforcement learning was proposed. Firstly, the simulation environment of manipulator was built, which includes the two-link manipulator, target and obstacle. Then, according to the target setting, state variables as well as reward and punishment mechanism of the environment model, three kinds of deep reinforcement learning models were established for training. Finally, the motion control of the two-link manipulator was realized. After comparing and analyzing the three proposed models, Deep Deterministic Policy Gradient (DDPG) algorithm was selected for further research to improve its applicability, so as to shorten the debugging time of the manipulator model, and avoided the obstacle to reach the target smoothly. Experimental results show that, the proposed deep reinforcement learning method can effectively control the motion of two-link manipulator, the improved DDPG algorithm control model has the convergence speed increased by two times and the stability after convergence enhances. Compared with the traditional control method, the proposed deep reinforcement learning control method has higher efficiency and stronger applicability.
    Reference | Related Articles | Metrics
    Remote sensing image dehazing method based on cascaded generative adversarial network
    SUN Xiao, XU Jindong
    Journal of Computer Applications    2021, 41 (8): 2440-2444.   DOI: 10.11772/j.issn.1001-9081.2020101563
    Abstract206)      PDF (2363KB)(340)       Save
    Dehazing algorithms based on image training pairs are difficult to deal with the problems of insufficient training sample pairs in remote sensing images, and have the model with weak generalization ability, therefore, a remote sensing image dehazing method based on cascaded Generative Adversarial Network (GAN) was proposed. In order to solve the missing of paired remote sensing datasets, U-Net GAN (UGAN) learning haze generation and Pixel Attention GAN (PAGAN) learning dehazing were proposed. In the proposed method, UGAN was used to learn how to add haze to the haze-free remote sensing images with the details of the images retained by using unpaired clear and haze image sets, and then was used to guide the PAGAN to learn how to correctly dehazing such images. To reduce the discrepancy between the synthetic haze remote sensing images and the dehazing remote sensing images, the self-attention mechanism was added to PAGAN. By the generator, the high-resolution detail features were generated by using cues from all feature locations in the low-resolution image. By the discriminator, the detail features in distant parts of the images were checked whether they are consistent with each other. Compared with the dehazing methods such as Feature Fusion Attention Network (FFANet), Gated Context Aggregation Network (GCANet) and Dark Channel Prior (DCP), this cascaded GAN method does not require a large number of paired data to train the network repeatedly. Experimental results show this method can remove haze and thin cloud effectively, and is better than the comparison methods on both visual effect and quantitative indices.
    Reference | Related Articles | Metrics
    Derivative-free few-shot learning based performance optimization method of pre-trained models with convolution structure
    Yaming LI, Kai XING, Hongwu DENG, Zhiyong WANG, Xuan HU
    Journal of Computer Applications    2022, 42 (2): 365-374.   DOI: 10.11772/j.issn.1001-9081.2021020230
    Abstract203)   HTML41)    PDF (841KB)(286)       Save

    Deep learning model with convolution structure has poor generalization performance in few-shot learning scenarios. Therefore, with AlexNet and ResNet as examples, a derivative-free few-shot learning based performance optimization method of convolution structured pre-trained models was proposed. Firstly, the sample data were modulated to generate the series data from the non-series data based on causal intervention, and the pre-trained model was pruned directly based on the co-integration test from the perspective of data distribution stability. Then, based on Capital Asset Pricing Model (CAPM) and optimal transmission theory, in the intermediate output process of the pre-trained model, the forward learning without gradient propagation was carried out, and a new structure was constructed, thereby generating the representation vectors with clear inter-class distinguishability in the distribution space. Finally, the generated effective features were adaptively weighted based on the self-attention mechanism, and the features were aggregated in the fully connected layer to generate the embedding vectors with weak correlation. Experimental results indicate that the proposed method can increase the Top-1 accuracies of the AlexNet and ResNet convolution structured pre-trained models on 100 classes of images in ImageNet 2012 dataset from 58.82%, 78.51% to 68.50%, 85.72%, respectively. Therefore, the proposed method can effectively improve the performance of convolution structured pre-trained models based on few-shot training data.

    Table and Figures | Reference | Related Articles | Metrics
    News recommendation model with deep feature fusion injecting attention mechanism
    Yuxi LIU, Yuqi LIU, Zonglin ZHANG, Zhihua WEI, Ran MIAO
    Journal of Computer Applications    2022, 42 (2): 426-432.   DOI: 10.11772/j.issn.1001-9081.2021050907
    Abstract203)   HTML32)    PDF (755KB)(143)       Save

    When mining news features and user features, the existing news recommendation models often lack comprehensiveness since they often fail to consider the relationship between the browsed news, the change of time series, and the importance of different news to users. At the same time, the existing models also have shortcomings in more fine-grained content feature mining. Therefore, a news recommendation model with deep feature fusion injecting attention mechanism was constructed, which can comprehensively and non-redundantly conduct user characterization and extract the features of more fine-grained news fragments. Firstly, a deep learning-based method was used to deeply extract the feature matrix of news text through the Convolutional Neural Network (CNN) injecting attention mechanism. By adding time series prediction to the news that users had browsed and injecting multi-head self-attention mechanism, the interest characteristics of users were extracted. Finally, a real Chinese dataset and English dataset were used to carry out experiments with convergence time, Mean Reciprocal Rank (MRR) and normalized Discounted Cumulative Gain (nDCG) as indicators. Compared with Neural news Recommendation with Multi-head Self-attention (NRMS) and other models, on the Chinese dataset, the proposed model has the average improvement rate of nDCG from -0.22% to 4.91% and MRR from -0.82% to 3.48%. Compared with the only model with negative improvement rate, the proposed model has the convergence time reduced by 7.63%. on the English dataset, the proposed model has the improvement rates reached 0.07% to 1.75% and 0.03% to 1.30% respectively on nDCG and MRR; At the same time this model always has fast convergence speed. Results of ablation experiments show that adding attention mechanism and time series prediction module is effective.

    Table and Figures | Reference | Related Articles | Metrics
    Feature construction and preliminary analysis of uncertainty for meta-learning
    Yan LI, Jie GUO, Bin FAN
    Journal of Computer Applications    2022, 42 (2): 343-348.   DOI: 10.11772/j.issn.1001-9081.2021071198
    Abstract202)   HTML62)    PDF (483KB)(147)       Save

    Meta-learning is the learning process of applying machine learning methods (meta-algorithms) to seek the mapping between features of a problem (meta-features) and relative performance measures of the algorithm, thereby forming the learning process of meta-knowledge. How to construct and extract meta-features is an important research content. Concerning the problem that most of meta-features used in the existing related researches are statistical features of data, uncertainty modeling was proposed and the impact of uncertainty on learning system was studied. Based on inconsistency of data, complexity of boundary, uncertainty of model output, linear capability to be classified, degree of attribute overlap, and uncertainty of feature space, six kinds of uncertainty meta-features were established for data or models. At the same time,the uncertainty size of the learning problem itself was measured from different perspectives, and specific definitions were given. The correlations between these meta-features were analyzed on artificial datasets and real datasets of a large number of classification problems, and multiple classification algorithms such as K-Nearest Neighbor (KNN) were used to conduct a preliminary analysis of the correlation between meta-features and test accuracy. Results show that the average degree of correlation is about 0.8, indicating that these meta-features have a significant impact on learning performance.

    Table and Figures | Reference | Related Articles | Metrics
    Disambiguation method of multi-feature fusion based on HowNet sememe and Word2vec word embedding representation
    WANG Wei, ZHAO Erping, CUI Zhiyuan, SUN Hao
    Journal of Computer Applications    2021, 41 (8): 2193-2198.   DOI: 10.11772/j.issn.1001-9081.2020101625
    Abstract200)      PDF (1018KB)(356)       Save
    Aiming at the problems that the low-frequency words expressed by the existing word vectors are of poor quality, the semantic information expressed by them is easy to be confused, and the existing disambiguation models cannot distinguish polysemous words accurately, a multi-feature fusion disambiguation method based on word vector fusion was proposed. In the method, the word vectors expressed by HowNet sememes and the word vectors generated by Word2vec (Word to vector) were fused to complement the polysemous information of words and improve the expression quality of low-frequency words. Firstly, the cosine similarity between the entity to be disambiguated and the candidate entity was calculated to obtain the similarity between them. After that, the clustering algorithm and HowNet knowledge base were used to obtain entity category feature similarity. Then, the improved Latent Dirichlet Allocation (LDA) topic model was used to extract the topic keywords to calculate the similarity of entity topic feature similarity. Finally, the word sense disambiguation of polysemous words was realized by weighted fusion of the above three types of feature similarities. Experimental results conducted on the test set of the Tibet animal husbandry field show that the accuracy of the proposed method (90.1%) is 7.6 percentage points higher than that of typical graph model disambiguation method.
    Reference | Related Articles | Metrics
    Ordinal decision tree algorithm based on fuzzy advantage complementary mutual information
    WANG Yahui, QIAN Yuhua, LIU Guoqing
    Journal of Computer Applications    2021, 41 (10): 2785-2792.   DOI: 10.11772/j.issn.1001-9081.2020122006
    Abstract198)      PDF (1344KB)(304)       Save
    When the traditional decision tree algorithm is applied to the ordinal classification task, there are two problems:the traditional decision tree algorithm does not introduce the order relation, so it cannot learn and extract the order structure of the dataset; in real life, there is a lot of fuzzy but not exact knowledge, however the traditional decision tree algorithm cannot deal with the data with fuzzy attribute value. To solve these problems, an ordinal decision tree algorithm based on fuzzy advantage complementary mutual information was proposed. Firstly, the dominant set was used to represent the order relations in the data, and the fuzzy set was introduced to calculate the dominant set for forming a fuzzy dominant set. The fuzzy dominant set was able to not only reflect the order information in the data, but also obtain the inaccurate knowledge automatically. Then, the complementary mutual information was generalized on the basis of fuzzy dominant set, and the fuzzy advantage complementary mutual information was proposed. Finally, the fuzzy advantage complementary mutual information was used as a heuristic method, and an decision tree algorithm based on fuzzy advantage complementary mutual information was designed. Experimental results on 5 synthetic datasets and 9 real datasets show that, the proposed algorithm has less classification errors compared with the classical decision tree algorithm on the ordinal classification tasks.
    Reference | Related Articles | Metrics
    Structure-fuzzy multi-class support vector machine algorithm based on pinball loss
    Kai LI, Jie LI
    Journal of Computer Applications    2021, 41 (11): 3104-3112.   DOI: 10.11772/j.issn.1001-9081.2021010062
    Abstract194)   HTML32)    PDF (816KB)(116)       Save

    The Multi-Class Support Vector Machine (MSVM) has the defects such as strong sensitivity to noise, instability to resampling data and lower generalization performance. In order to solve the problems, the pinball loss function, sample fuzzy membership degree and sample structural information were introduced into the Simplified Multi-Class Support Vector Machine (SimMSVM) algorithm, and a structure-fuzzy multi-class support vector machine algorithm based on pinball loss, namely Pin-SFSimMSVM, was proposed. Experimental results on synthetic datasets, UCI datasets and UCI datasets adding different proportions of noise show that, the accuracy of the proposed Pin-SFSimMSVM algorithm is increased by 0~5.25 percentage points compared with that of SimMSVM algorithm. The results also show that the proposed algorithm not only has the advantages of avoiding indivisible areas of multi-class data and fast calculation speed, but also has good insensitivity to noise and stability to resampling data. At the same time, the proposed algorithm considers the fact that different data samples play different roles in classification and the important prior knowledge contained in the data, so that the classifier training is more accurate.

    Table and Figures | Reference | Related Articles | Metrics
    Deep network embedding method based on community optimization
    LI Yafang, LIANG Ye, FENG Weiwei, ZU Baokai, KANG Yujian
    Journal of Computer Applications    2021, 41 (7): 1956-1963.   DOI: 10.11772/j.issn.1001-9081.2020081193
    Abstract190)      PDF (1616KB)(289)       Save
    With the rapid development of technologies such as modern network communication and social media, the networked big data is difficult to be applied due to the lack of efficient and available node representation. Network representation learning is widely concerned by transforming high-dimensional sparse network data into low-dimensional, compact and easy-to-apply node representation. However, the existing network embedding methods obtain the low-dimensional feature vectors of nodes and then use them as the inputs for other applications (such as node classification, community discovery, link prediction and visualization) for further analysis, without building models for specific applications, which makes it difficult to achieve satisfactory results. For the specific application of network community discovery, a deep auto-encoder clustering model that combines community structure optimization for low-dimensional feature representation of nodes was proposed, namely Community-Aware Deep Network Embedding (CADNE). Firstly, based on the deep auto-encoder model, the node low-dimensional representation was learned by maintaining the topological characteristics of the local and global links of the network, and then the low-dimensional representation of the nodes was further optimized by using the network clustering structure. In this method, the low-dimensional representations of the nodes and the indicator vectors of the communities that the nodes belong to were learnt at the same time, so that the low-dimensional representation of the nodes can not only maintain the topological characteristics of the original network structure, but also maintain the clustering characteristics of the nodes. Comparing with the existing classical network embedding methods, the results show that CADNE achieves the best clustering results on Citeseer and Cora datasets, and improves the accuracy by up to 0.525 on 20NewsGroup. In classification task, CADNE performs the best on Blogcatalog and Citeseer datasets and the performance on Blogcatalog is improved by up to 0.512 with 20% training samples. In the visualization comparison, CADNE molel can get a low-dimensional representation of nodes with clearer class boundary, which verifies that the proposed method has better low-dimensional representation ability of nodes.
    Reference | Related Articles | Metrics
2022 Vol.42 No.3

Current Issue
Archive
Honorary Editor-in-Chief: ZHANG Jingzhong
Editor-in-Chief: XU Zongben
Associate Editor: SHEN Hengtao XIA Zhaohui
Domestic Post Distribution Code: 62-110
Foreign Distribution Code: M4616
Address:
No. 9, 4th Section of South Renmin Road, Chengdu 610041, China
Tel: 028-85224283-803
  028-85222239-803
Website: www.joca.cn
E-mail: bjb@joca.cn
WeChat
Join CCF