Computer Science

IntervalSketch:Approximate Statistical Method for Interval Items in Data Stream

CHEN Xinyang, CHEN Hanze, ZHOU Jiasheng, HUANG Jiaqing, YU Jiashuo, ZHU Longlong, ZHANG Dong

Computer Science. 2024, 51 (4): 4-10. doi:10.11896/jsjkx.231000226

Abstract

PDF(2197KB) ( 278 )

References | Related Articles | Metrics

The proportion of streaming databases is gradually increasing,and extracting the required information in the data streams of streaming databases is an important task.In this paper,we study interval items which refer to pairs of elements arriving with a fixed interval,and apply them to network scenarios.It is the first work to define and count interval items in data streams.To efficiently count the top-K interval items,IntervalSketch is proposed.IntervalSketch firstly chunks the data stream based on simulated annealing to accelerate the statistical speed,secondly,it uses Sketch to store the interval items,and lastly reduces the memory of storing the interval items in Sketch through the feature grouping storage strategy,which enhances the accuracy of counting the interval items.Extensive comparative experiments are carried out on two real datasets.Experimental results show that IntervalSketch significantly outperforms the baseline solution with the same memory,and the processing time is1/3～1/2 of the baseline solution,the average absolute error and the average relative error are1/3 of the baseline solution.

Large-scale Network Community Detection Algorithm Based on MapReduce

WANG Hancheng, DAI Haipeng, CHEN Zhipeng, CHEN Shusen, CHEN Guihai

Computer Science. 2024, 51 (4): 11-18. doi:10.11896/jsjkx.231100049

Abstract

PDF(2053KB) ( 322 )

References | Related Articles | Metrics

Community detection is a fundamental problem in the field of social network mining.With the rapid generation of massive data,traditional community detection algorithms are becoming increasingly difficult to handle large-scale social networks.Therefore,it is of great significance to design efficient community detection algorithms for large-scale networks.This paper proposes a new distributed algorithm based on MapReduce and k-center clustering.Firstly,the algorithm proposes the “friend circle coefficient” technique,which can measure the distance between nodes more accurately.Secondly,the algorithm proposes the “two-stage k-center clustering” technique,which incorporates node centrality heuristic information into the process of selecting center points and can significantly optimize the modularity of the results.Finally,the algorithm proposes a “community fusion method with modularity as the optimization goal” technique,which can automatically determine the number of communities in the network without prior knowledge.The evaluation results show that the proposed algorithm significantly outperforms the state-of-the-art community discovery algorithms in terms of modularity.For example,compared with the LPA algorithm,the proposed algorithm increases the modularity by an average of 9.19 times.

Data Quality Measurement Framework Research and Field Measurement Framework Construction

SONG Jinyu, CHEN Lianyong, CHEN Gang

Computer Science. 2024, 51 (4): 19-27. doi:10.11896/jsjkx.230400138

Abstract

PDF(1901KB) ( 192 )

References | Related Articles | Metrics

In order to activate the potential of data quality,a data quality measurement framework that takes into account information environment and technology realization is constructed,so as to improve the effectiveness of data mining and command decision.At first,the existing general and industrial data quality measurement frameworks are studied from the macro and micro levels,the data quality dimension clusters are obtained by clustering the data quality dimensions,and two types of characteristics of data quality dimension are extracted.Furthermore,the construction guidelines of data quality measurement framework for specific field are put forward.Finally,based on the requirements of data quality measurement in management field,the data quality measurement framework for management field(DQMFM) is constructed by combining construction guidelines.Besides,the data quality dimensions,measurement metrics and measurement methods of DQMFM are introduced.

Global Top-K Frequent Flow Measurement for Continuous Periods in Distributed Networks

MAO Chenyu, HUANG He, SUN Yu'e, DU Yang

Computer Science. 2024, 51 (4): 28-38. doi:10.11896/jsjkx.231000119

Abstract

PDF(4350KB) ( 106 )

References | Related Articles | Metrics

In distributed networks,the measurment of the top-K frequent flows is crucial for applications like resource allocation and security monitoring.Existing works on top-K frequent flow measurement have limitations such as being unsuitable for distributed network traffic measurement or only considering single time periods.To address these problems,this paper proposes a scheme for measuring global top-K frequent flows over continuous time periods in distributed networks.This involves deploying compact probabilistic data structures at distributed nodes to record network flow information.At the end of each time period,distributed nodes send necessary information to a central node,which aggregates this to identify the global top-K frequent flows from the start of measurement to the current time period.Considering that each flow may appear at one or multiple measurement nodes,different methods are used to reduce transmission overhead.For flows appearing at a single node,a method of transmitting segmented minimum values is used to obtain a threshold.Experiments show that this method reduces the transmission overhead by over 50% compared to full transmission.For flows appearing at multiple nodes,a multi-stage error-free processing method and a single-stage fast processing method are proposed,catering to scenarios that cannot tolerate errors and actual high-speed network traffic,respectively.Compared to using existing single-period methods in each time period,experimental performance of transmission overhead reduced by two orders of magnitude.Finally,a method using historical average increment information to reduce communication delay is also proposed,and experimental results show that it effectively reduces the average relative error of constraint information.

Multi-level Pruning Obfs4 Obfuscated Traffic Recognition Method Based on Partial Data

XU Chenhan, HUANG He, SUN Yu'e, DU Yang

Computer Science. 2024, 51 (4): 39-47. doi:10.11896/jsjkx.231000118

Abstract

PDF(2135KB) ( 113 )

References | Related Articles | Metrics

Obfs4 obfuscated traffic,carried by the anonymous communication network Tor,is often misused for illicit online acti-vities due to its strong anonymity.Consequently,the identification of Obfs4 obfuscated traffic plays a critical role in preventing cybercrime via the Tor network.Existing methods tend to focus on the analysis of Obfs4 traffic features,utilize machine learning or deep learning techniques for the precise identification of entire flow samples.However,in the realm of flow recognition,it often results in considerable time overhead.Recognition accuracy also decreases notably with the incorporation of inter-arrival timing(IAT) technology in Obfs4.In response,a multi-level pruning method for Obfs4 obfuscated traffic recognition based on partial data is proposed.This approach involves collecting only a small number of initial packets from each flow for several rounds of rapid filtering,and is specifically designed to enhance the efficiency and reliability of Obfs4 traffic identification by focusing on the IAT pattern.The approach breaks down the process into two key phases:a handshake phase and an encrypted communication phase.During the handshake phase,it thoroughly explores the underlying meanings in Obfs4 handshake packets,enabling quick filtering based on broad characteristics like randomness,timing,and length distribution.In the encrypted communication phase,it extracts features from the first packets of each flow and places greater importance on features related to IAT.Finally,fine-grained identification is accomplished using the XGBoost classification method.Experimental findings indicate that despite the implementation of IAT technology,leveraging the initial 30～50 data packets from the flow yields a 99% accuracy rate,with an average processing time per flow measured in milliseconds.

RBFRadar:Detecting Remarkable Burst Flows with Programmable Data Plane

WU Yanni, ZHOU Zhengyan, CHEN Hanze, ZHANG Dong

Computer Science. 2024, 51 (4): 48-55. doi:10.11896/jsjkx.231000213

Abstract

PDF(1979KB) ( 90 )

References | Related Articles | Metrics

Burst is a common and important traffic pattern in diverse network traffics.Since bursts may increase network latency and have a non-trivial impact on application performance,the efforts to detect,analyze and mitigate burst flows are meaningful for improving the performance and robustness of network.However,existing per-burst-based detection schemes face the limitations of significant bandwidth overheads and high user burdens.This paper proposes the detection of remarkable burst flows(RBFs) via observing and analyzing the characteristics of burst flows in various scenarios.The detection of RBFs reduces the bandwidth overheads.At the same time,such detection process avoids the requirements of intensive manual labor and expert experience,and mitigate the burdens of network operators.We propose RBFRadar,a Sketch-based RBF detection framework that supports RBF detection on programmable data plane,observing flow-level burstiness in a period.We prototype RBFRadar in PISA architecture with limited memory footprints and low time complexity.Experiments demonstrate that the F1-score of RBFRadar in RBF detection is 5.6 times to 23.4 times higher than that of existing schemes.Compared with per-burst detection,the bandwidth overhead could be reduced by 84.62% to 98.84%.

Performance Optimization of Complex Stencil in Weather Forecast Model WRF

DI Jianqiang, YUAN Liang, ZHANG Yunquan, ZHANG Sijia

Computer Science. 2024, 51 (4): 56-66. doi:10.11896/jsjkx.231000124

Abstract

PDF(2329KB) ( 95 )

References | Related Articles | Metrics

The weather research and forecasting model(WRF) is a widely used mesoscale numerical weather forecasting system that plays an important role in the fields of atmospheric research and meteorological operational forecasting.Stencil computation is a common nested loop pattern in scientific and engineering applications.WRF performs a large number of complex stencil computation on spatial grids to solve numerical equations of atmospheric dynamics and thermodynamics.The stencils in WRF are featured by multi-dimensionality,multi-variables,particularity of physical model boundaries,and complexity of physical and dynamic processes.This study analyzes the typical stencil pattern in WRF,identifies and abstracts the concept of “intermediate variable”,and implements three optimization schemes,namely,intermediate variable computation merging,intermediate variable dimensio-nality reduction storage,and intermediate variables extraction.The optimization schemes effectively improve the data locality,increase data reuse and spatial reuse rates,and reduces redundant computing and memory access overhead.The results show that the WRF 4.2 typical hotspot functions achieve significant performance improvements on both Intel CPU and Hygon CPU,with the highest speedup ratios of 21.3% and 17.8% respectively.

Transplantation and Optimization of Graph Matching Algorithm Based on Domestic DCUHeterogeneous Platform

HAO Meng, TIAN Xueyang, LU Gangzhao, LIU Yi, ZHANG Weizhe, HE Hui

Computer Science. 2024, 51 (4): 67-77. doi:10.11896/jsjkx.230800193

Abstract

PDF(3041KB) ( 120 )

References | Related Articles | Metrics

Subgraph matching is a basic graph algorithm that is widely used in various fields such as social networks and graph neural networks.As the scale of graph data grows,there is an increasing need for efficient subgraph matching algorithms.GENEVA is a GPU-based parallel subgraph matching algorithm.It uses the interval index graph storage structure and parallel matching optimization method to greatly reduce storage overhead and improve subgraph matching performance.However,due to the diffe-rence in the underlying hardware architecture and compilation environment of the platform,GENEVA cannot be directly applied to the domestic DCU platform.In order to solve this problem,this paper proposes GENEVA's transplantation and optimization scheme for domestic DCU.IO time consumption is the main performance bottleneck of GENEVA algorithm.This paper proposes three optimization strategies of page-locked memory,preloading,and scheduler to alleviate this bottleneck.Among them,page-locked memory technology avoids additional data transfer from pageable memory to temporary page-locked memory,and greatly reduces the time consumption of IO transfer on the DCU platform.The preloading technology overlaps IO data transmission with DCU kernel function computation to mask IO time consumption.The scheduler reduces redundant data transfer while satisfy preloading requirements.In this paper,Experiments are carried out on three real-world datasets of different sizes,and the results show that the algorithm performance is significantly improved after using the optimization strategies.On 92.6% of the test cases,the optimized GENEVA-HIP execution time on the Sugon DCU platform is less than that of the unported GENEVA on the GPU server.On a larger dataset,the execution time of the optimized Geneva-HIP algorithm on the DCU platform is reduced by 52.73% compared with the the pre-port GENEVA algorithm on the GPU server.

Auto-vectorization Cost Model Based on Instruction MKS

WANG Zhen, NIE Kai, HAN Lin

Computer Science. 2024, 51 (4): 78-85. doi:10.11896/jsjkx.230200024

Abstract

PDF(2431KB) ( 72 )

References | Related Articles | Metrics

The auto-vectorization cost model is an important component of compiler's auto-vectorization optimization.Its role is to evaluate whether the code can achieve performance improvement after applying vectorization transformation.When the cost model is inaccurate,the compiler will apply vectorization transformation with negative benefit,thus reducing the execution efficiency of the program.Aiming at the inaccuracy of the default cost model of GCC compiler,based on Intel Xeon Silver 4214R CPU,an auto-vectorization cost model based on instruction MKS is proposed.The model fully considers the machine mode,operation type and operation intensity of instructions,and uses gradient descent algorithm to automatically search the approximate cost of different instruction types.Single-thread tests are carried out on SPEC2006 and SPEC2017.Experimental results show that the model can reduce the error of benefit estimation.Compared with the vector program generated by the default cost model,the GCC compiler,after adding the MKS cost model,achieves a maximum speedup of 4.72% on the SPEC2006 benchmark and 7.08% on the SPEC2017 benchmark.

Floating-point Expression Precision Optimization Method Based on Multi-type Calculation
Rewriting

HAO Jiangwei, YANG Hongru, XIA Yuanyuan, LIU Yi, XU Jinchen , PANG Jianmin

Computer Science. 2024, 51 (4): 86-94. doi:10.11896/jsjkx.221200072

Abstract

PDF(2371KB) ( 73 )

References | Related Articles | Metrics

Expression rewriting is an emerging method in the field of precision optimization.Its core idea is to transform an expression into a semantically equivalent expression without changing its precision representation to improve precision.However,given the large number of transformation rules and the huge transformation space,the problem of the rewriting method is how to choose an appropriate transformation strategy.In response to the above problems,this paper proposes a precision optimization method for floating-point expressions based on multi-type calculation rewriting,which supports expressions including mathematical function calculations and four arithmetic calculations,and implements an expression rewriting tool exprAuto.Unlike other precision optimization tools that focus on replacing sub-expressions,exprAuto pays more attention to transform the order of expression operations.After the expression is simplified and mathematically transformed,exprAuto obtains different calculation orders through polynomial transformation,and tries to improve the precision by reducing the number of operations.Finally,exprAuto generates an equivalent set of expressions with different calculation orders and selects the final precision optimization result through sorting screening and error detection.In this paper,41 expressions from the FPBench standard set and 18 approximate polynomials of common mathematical functions are selected as test cases.After exprAuto optimization,the maximum error is reduced by 45.92% and the average error is reduced by 34.98% compared to the original expression.For 18 approximate polynomials,the maximum error is reduced by 58.35%,and the average error is reduced by 43.73%.Experimental results show that exprAuto can effectively improve the precision of expressions,especially polynomials.

Review of Node Classification Methods Based on Graph Convolutional Neural Networks

ZHANG Liying, SUN Haihang, SUN Yufa , SHI Bingbo

Computer Science. 2024, 51 (4): 95-105. doi:10.11896/jsjkx.230600071

Abstract

PDF(1683KB) ( 162 )

References | Related Articles | Metrics

Node classification is one of the important research tasks in graph field.In recent years,with the continuous deepening of research on graph convolutional neural network,significant progress has been made in the research and application of node classification based on graph convolutional neural networks.Graph convolutional neural networks are kind of graph neural network method based on convolution.It can handle graph data and have the advantages of convolutional neural networks,and have become the most active branch of graph node classification research.This paper first introduces the related concepts of graph,the definition of node classification and commonly used graph datasets.Then,it reviews two classic graph convolutional neural networks,spectral domain and spatial domain graph convolutional neural networks,and discusses the challenges of using graph con-volutional neural networks to study node classification.Next,it analyzes the research progress and unresolved issues of graph convolutional neural networks in node classification tasks from the perspectives of model and data.Finally,this paper gives insights into the research direction on node classification based on graph convolutional neural networks.

Node Influence Ranking Model Based on Transformer

XI Ying, WU Xuemeng, CUI Xiaohui

Computer Science. 2024, 51 (4): 106-116. doi:10.11896/jsjkx.230300110

Abstract

PDF(3477KB) ( 87 )

References | Related Articles | Metrics

Node influence ranking is a key topic in complex networks,and plays an important role in identifying key nodes and measuring node influence.There has been much research exploring node influence based on complex networks,with deep learning shows great potential.However,existing convolutional neural networks(CNNs) and graph neural networks(GNNs) are often based on fixed dimensional features as input and cannot effectively distinguish between neighboring nodes,making them unsuitable for diverse complex networks.In order to solve these problems,a simple and effective node influence ranking model is proposed in this paper.In this model,the input sequence of nodes contains information about the nodes themselves and their neighbors,and the length of the input sequence can be dynamically adjusted according to the network to ensure that the model obtains sufficient information about the nodes.The model also uses the self-attention mechanism to enable nodes to efficiently aggregate information about their neighbors in the input sequence,thus identifying the influence of nodes.Experiments are conducted on 12 real network datasets to verify the effectiveness of the model against seven existing methods using multi-dimensional evaluation criteria.Experimental results show that the model can identify the influence of nodes in complex networks more effectively.

Graph Sampling Algorithm Based on Representative Node Expansion to Maintain CommunityStructure

HONG Yu, CHEN Hongchang, ZHANG Jianpeng, HUANG Ruiyang , LI Shaomei

Computer Science. 2024, 51 (4): 117-123. doi:10.11896/jsjkx.230100018

Abstract

PDF(1789KB) ( 61 )

References | Related Articles | Metrics

Graph sampling is widely used in real life as a method to simplify large-scale graphs and retain specified properties.However,most of the current research focuses on preserving node-level properties,such as degree distribution,while ignoring more important information such as the community structure of graphs.To solve this problem,a graph sampling algorithm is proposed to maintain the community structure.The algorithm is divided into two steps.The first step is to initialize the community representative points,and the node importance is calculated according to the proposed node importance calculation formula,and then the representative nodes of each community are selected.The second step is to expand the community structure.For each community,it selects the node that can introduce the least additional neighbors to join the community until the upper limit of the community node is reached.Comparative experiments are conducted on a number of real data sets,and multiple evaluation indicators are adopted to evaluate the experimental results.Experimental results show that the proposed sampling algorithm can well maintain the overall community structure,and provides a feasible solution for sampling community structure of large-scale graphs.

Semi-supervised Classification of Data Stream with Concept Drift Based on Clustering Model Reuse

KANG Wei, LI Lihui, WEN Yimin

Computer Science. 2024, 51 (4): 124-131. doi:10.11896/jsjkx.230300023

Abstract

PDF(1987KB) ( 72 )

References | Related Articles | Metrics

Semi-supervised classification of data stream with concept drift poses challenges to classifier training,classifier adaption for new concept,and concept drifting detection,for only some or even very few instances are labeled.In the existing semi-supervised clustering classification algorithms,only the clustering model in the classifier pool is updated incrementally,and the historical clustering model cannot be reused effectively.Therefore,this paper proposes a new cluster-based model reuse semi-supervised classification algorithm,CDCMR.First,the data stream comes in the form of data chunks.After classifying the data chunks,a clustering model with adaptive determination of the number of clusters is trained.Secondly,multiple history classifiers are selected by calculating the similarity between each history classifier in the classifier pool and the clustering model.Thirdly,the selected history classifier is reused with the current data chunk and integrated with the cluster model.Then,the classifier pool is divided into old and new replacement and diversity maximization classifier pool for updating.Finally,the samples of the next data chunk are ensemble classification.Experimental results on several artificial and real data sets show that the algorithm can effectively adapt to concept drift,which is significantly improved compared with the existing methods.

Benchmarking and Analysis for Graph Neural Network Node Classification Task

ZHANG Tao, LIAO Bin, YU Jiong, LI Ming, SUN Ruina

Computer Science. 2024, 51 (4): 132-150. doi:10.11896/jsjkx.230200084

Abstract

PDF(11589KB) ( 96 )

References | Related Articles | Metrics

In contrast with previous graph embedding algorithms,the graph neural network model performs tasks such as node classification more effectively because it can better coordinate the learning of hidden node features with the classification target due to its end-to-end model architecture in the training process.However,the experimental comparison stage of existing graph neural models frequently suffers from problems such as specific types of experimental datasets,insufficient dataset sample size,irregular splitting of the train and test sets,limited scale and scope of comparison models,homogeneous performance evaluation metrics,and lack of comparative analysis for model's training time consumption.To this end,in order to provide decision guidelines for GNN model selection in real business scenarios,a total of 20 datasets from various domains(citation networks,social networks,collaboration networks,etc.),including cora,citeseer,pubmed,deezer,etc.,are chosen to conduct a comprehensive and equitable benchmark evaluation of node classification tasks on 17 mainstream graph neural network models,including FastGCN,PPNP,ChebyNet,DAGNN,etc.,on performance evaluation metrics including accuracy,precision,recall,F-score value,and model training time.The benchmarking experiments revealed that,on the one hand,the factors that affect the speed of model training are node attribute dimension,graph node size and graph edge size in turn;on the other hand,there is no winner-take-all model,that is,there is no model that performs well across all benchmark datasets,especially in a fair benchmarking configuration,the model with simple structure has better performance than the complex GNN models.

Traffic Flow Prediction Model Based on Dual Prior-adaptive Graph Neural ODE Network

YUAN Rong, PENG Lilan, LI Tianrui, LI Chongshou

Computer Science. 2024, 51 (4): 151-157. doi:10.11896/jsjkx.230100066

Abstract

PDF(2049KB) ( 77 )

References | Related Articles | Metrics

Accurate traffic flow prediction is an indispensable part of intelligent transportation system.In recent years,graph neural networks have generated effective results in traffic flow prediction tasks.However,the information transfer of graph neural network is discontinuous latent state propagation,and there is an over-smoothing problem as the number of network layers increases,which limits the ability of the model to capture the spatial dependencies of distant nodes.At the same time,when representing the spatial relationship of the road network,most of the existing methods only use the predefined graph constructed by prior knowledge or the adaptive graph constructed only by the road network conditions,ignoring the combination of those two graphs.Aiming at solving the above problems,this paper proposes a traffic flow prediction model based on a dual prior adaptive graph neural ordinary differential equation.Temporal convolutional network are utilized to capture the temporal correlation of sequences,a priori adaptive graph fusion module is used to represent the road network,and complex spatio-temporal features are propagated in a continuous manner through tensor multiplication-based nerual ODEs.Finally,experiments are carried out on four public data sets of highway traffic in California,USA.Experimental results show that the prediction performance of the model is better than that of the existing ten methods.

Urban Electricity Load Forecasting Method Based on Discrepancy Compensation and Short-termSampling Contrastive Loss

CHEN Runhuan, DAI Hua, ZHENG Guineng, LI Hui , YANG Geng

Computer Science. 2024, 51 (4): 158-164. doi:10.11896/jsjkx.230100089

Abstract

PDF(3180KB) ( 62 )

References | Related Articles | Metrics

Urban power load forecasting is an important content of urban smart grid planning and scheduling.However,the pro-blem of data imbalance in urban power load forecasting poses a great challenge to urban power load forecasting.Traditional single-model-based methods can hardly solve the problem of data imbalance.The existing multi-model-based forecasting methods split the datasets into multiple sub-datasets according to the electricity load profiles,and then build multiple forecasting models for forecasting,which can solve the data imbalance problem to a certain extent,but there are problems such as high model construction cost and separation of the common electricity distribution characteristics among different distribution profiles.Based on this,this paper proposes a lighten urban electric load forecasting model(Lighten-DCSC-LSTM).It is constructed by introducing the idea of discrepancy compensation and short-term sampling contrastive loss on the basis of long and short-term memory networks,while building a shared feature extraction layer to reduce the model construction cost.Among them,the idea of discrepancy compensation compensates the prediction results of the main sequence prediction module by learning the differences between different power load distribution samples,and the short-term sampling contrastive loss regularizes the training of the model by the contrastive learning loss of the dynamic class center.To verify the performance of the proposed model,parameter tuning and comparison experiments are conducted.The results of the comparison experiments show that the model achieves good perfor-mance in the task of forecasting electricity loads.

Data Completion and Prediction of Street Parking Spaces Based on Transformer

LIN Binwei, YU Zhiyong, HUANG Fangwan, GUO Xianwei

Computer Science. 2024, 51 (4): 165-173. doi:10.11896/jsjkx.221200171

Abstract

PDF(2892KB) ( 83 )

References | Related Articles | Metrics

With the continuous growth of the number of cars in cities,the difficulty of parking on the street has become a hot issue.The key to solve the street parking problem is to accurately predict the future parking space information of the street.CrowdSensing is a low-cost and cost-effective way of sensing parking space by installing sonar on vehicles.However,the parking space data sensed in this way has high sparsity in time,and the traditional model cannot be directly used for prediction.To solve this problem,a transformer-based parking space sequence completion and prediction network is proposed.This network generates the memory of the missing parking space sequence through the encoder,and then the decoder completes the missing part of the parking space sequence in the way of auto-regression,and predicts the future parking space information.Experimental results show that the proposed method is better than a series of traditional machine learning and deep learning methods in the completion and prediction of two highly missing street parking space data sets.

Study on Manufacturing Company Automated Chart Analysis Method Based on Natural LanguageGeneration

WANG Xu, LIU Changhong, LI Shengchun, LIU Shuang, ZHAO Kangting, CHEN Liang

Computer Science. 2024, 51 (4): 174-181. doi:10.11896/jsjkx.230400031

Abstract

PDF(3038KB) ( 73 )

References | Related Articles | Metrics

With the wave of digital transformation,manufacturing enterprises produce a large number of chart data every day.Traditional chart analysis methods are difficult to analyze chart data efficiently and accurately.Automated chart analysis methods have become an important means of chart analysis.In order to solve the problem that the automatic chart analysis method is difficult to meet the specific needs in practical application,an automatic chart analysis method of manufacturing enterprises based on natural language generation is proposed.This method analyzes the chart data based on LSTM,and in order to solve the problem of misleading LSTM by redundant data in the analysis process,a discriminator layer is added after the embedding layer to enable LSTM to perform more targeted semantic understanding and text prediction according to the type of chart.Aiming at the problem of poor quality of description sentences generated in the process of diagram analysis,a random cluster sampling strategy is proposed to improve the quality of diagram analysis by referring to beam search and random sampling strategy,and knowledge distillation method is introduced to optimize LSTM to further improve the quality of description text.Experiments show that this method improves the text quality by 8.9% compared with LSTM.In order to apply the method in practice,an automatic chart analysis system for manufacturing enterprises is designed and developed,and the method is introduced as a chart analysis tool.Experimental results show that the application of this method can improve the quality and efficiency of chart analysis in manufacturing enterprises.

Study on Time-varying Brain State Based on fMRI Data－A Review

LIN Qiye, XIA Jianan, ZHOU Xuezhong

Computer Science. 2024, 51 (4): 182-192. doi:10.11896/jsjkx.230700059

Abstract

PDF(2651KB) ( 96 )

References | Related Articles | Metrics

Functional magnetic resonance imaging(fMRI) has been widely applied in the study of human brain activity.Recently,the use of brain states to investigate brain dynamics has attracted extensive attention from researchers.Previous reviews on brain states typically compare and summarize from the perspective of state definition methods,neglecting the inconsistency in under-lying data formats,which may results in diverse interpretations of brain states.Furthermore,these reviews also lack discussions on the analytical approaches for brain states.Here,we review various methods for defining brain states based on different data formats,provide an overview of different approaches for analyzing brain dynamics based on brain states,and summarize typical research methods in the application of brain states to cognition,psychiatric disorders,physiological states,and other aspects.Fina-lly,we find similarities between the definition of brain meta-states and feature extraction in deep learning.Therefore,we believe that deep learning is a promising approach for studying brain states.

Review of Vision-based Neural Network 3D Dynamic Gesture Recognition Methods

WANG Ruiping, WU Shihong, ZHANG Meihang, WANG Xiaoping

Computer Science. 2024, 51 (4): 193-208. doi:10.11896/jsjkx.230200205

Abstract

PDF(7133KB) ( 113 )

References | Related Articles | Metrics

Dynamic gesture recognition,as an important means of human-computer interaction,has received widespread attention.Among them,the visual-based recognition method has become the preferred choice for the new generation of human-computer interaction due to its convenience and low cost.Centered on artificial neural networks,this paper reviews the research progress of visual-based gesture recognition methods,analyzes the development status of different types of artificial neural networks in gesture recognition,investigates and summarizes the types and characteristics of data to be recognized and training datasets.In addition,through performance comparison experiments,different types of artificial neural networks are objectively evaluated,and the results are analyzed.Finally,based on the summary of the research content,the challenges and problems faced in this field are elaborated,and the development trend of dynamic gesture recognition technology is prospected.

Metal Surface Defect Detection Method Based on Dual-stream YOLOv4

XU Hao, LI Fengrun, LU Lu

Computer Science. 2024, 51 (4): 209-216. doi:10.11896/jsjkx.230100141

Abstract

PDF(3452KB) ( 148 )

References | Related Articles | Metrics

Currently,many researchers use deep learning for surface defect detection.However,most of these studies follow the mainstream object detection algorithm and focus on high-level semantic features while neglecting the importance of low-level semantic information(color,shape) for surface defect detection,resulting in unsatisfactory defect detection effect.To address this issue,a metal surface defect detection network called the dual-stream YOLOv4 network is proposed.The backbone network is split into two branches,with inputs consisting of high-resolution and low-resolution images.The shallow branch is responsible for extracting low-level features from the high-resolution image,while the deep branch is responsible for extracting high-level features from the low-resolution image.The model's total parameter volume is reduced by cutting down the number of layers and channels in both branches.To enhance the low-level semantic features,a tree-structured multi-scale feature fusion method(TMFF) is proposed,and a feature fusion module with a polarized self-attention mechanism and spatial pyramid pooling(FFM-PSASPP) is designed and applied to the TMFF.The algorithm's map@50 results on the test sets of the Northeastern University hot-rolled strip surface defect dataset(NEU-DET),the metal surface defect dataset(GC10-DET),and the enaiter rice cooker inner pot defect dataset are 0.80,0.66,and 0.57,respectively.Compared to most mainstream object detection algorithms used for defect detection,there is an improvement,and the model's parameter volume is only half that of the original YOLOv4,with a speed close to YOLOv4,making it suitable for practical use.

Video and Image Salient Object Detection Based on Multi-task Learning

LIU Zeyu, LIU Jianwei

Computer Science. 2024, 51 (4): 217-228. doi:10.11896/jsjkx.231000051

Abstract

PDF(5070KB) ( 116 )

References | Related Articles | Metrics

Salient object detection(SOD) can quickly identify high-value salient objects in complex scenes,which simulates human attention and lays the foundation for further vision understanding tasks.Currently,the mainstream methods for image-based salient object detection are usually trained on DUTS-TR dataset,while video-based salient object detection(VSOD) methods are trained on DAVIS,DAVSOD,and DUTS-TR datasets.Because image and video salient object detection tasks have general and specific characteristics,independent models need to be deployed for separate training,which greatly increases computational resources and training time.Current research typically focuses on independent solution for a single task.However,a unified method for both image and video salient object detection is lack of research.To address on aforementioned issues,this paper proposes a multi-task learning-based method for image and video salient object detection,aiming to build a universal framework which simultaneously adapts to both tasks with a single training process,and further bridges the performance gaps between image and video salient object detection models.Qualitative and quantitative experimental results on 12 datasets show that the proposed method can not only adapt to both tasks,but also achieve better detection results than single-task models.

Algorithm of Stereo Matching Based on GAANET

SONG Hao, MAO Kuanmin, ZHU Zhou

Computer Science. 2024, 51 (4): 229-235. doi:10.11896/jsjkx.230100137

Abstract

PDF(2825KB) ( 57 )

References | Related Articles | Metrics

End-to-end stereo matching algorithms have become increasingly popular in stereo matching tasks due to their advantages in computational time and matching accuracy.However,feature extraction in such algorithms can result in redundant features,information loss,and insufficient multi-scale feature fusion,thereby increasing computational complexity and decreasing matching accuracy.To address these challenges,an improved ghost adaptive aggregation network(GAANET) is proposed based on the adaptive aggregation network(AANET),and its feature extraction module is improved to make it more suitable for stereo matching tasks.Multi-scale features are extracted in the G-Ghost phase,and partial features are generated through low-cost ope-rations to reduce feature redundancy and preserve shallow features.An efficient channel attention mechanism is implemented to allocate weights to each channel,and an improved feature pyramid structure is introduced to mitigate channel information loss in traditional pyramids and optimize feature fusion,thus enhancing information supplement for features across scales.The proposed GAANET model is trained and evaluated on the SceneFlow,KITTI2015,and KITTI2012 datasets.Experimental resultsdemons-trate that GAANET outperforms the baseline method,with accuracy improvements of 0.92%,0.25%,and 0.20%,respectively,while reducing parameter volume by 13.75% and computational complexity by 4.8%.

Human Action Recognition Algorithm Based on Adaptive Shifted Graph Convolutional Neural
Network with 3D Skeleton Similarity

YAN Wenjie, YIN Yiying

Computer Science. 2024, 51 (4): 236-242. doi:10.11896/jsjkx.221200120

Abstract

PDF(2583KB) ( 61 )

References | Related Articles | Metrics

Graph convolutional neural network(GCN) has achieved good results in the field of human action recognition based on 3D skeleton.However,in most of the existing GCN methods,the construction of the behavior diagram is based on the manual setting of the physical structure of the human body.In the training stage,each graph node can only establish the connection accor-ding to the manual setting,which cannot perceive new connections between bone nodes during action,resulting in the unreasonable and inflexible topology of the graph.The shifted graph convolutional neural network(Shift-GCN) makes the receptive field more flexible by changing its structure,and achieves satisfied results in the global shift angle.In order to tackle the above pro-blems of graph structure,an adaptive shift graph convolutional neural network(AS-GCN) is proposed to make up for the above shortcomings.AS-GCN draws on the idea of shifted graph convolutional neural network,and proposes to use the characteristics of each human action to guide the graph network to perform shift operation,so as to select the nodes that need to expand the receptive field as accurately as possible.On the general skeleton-based action recognition dataset NTU-RGBD,the AS-GCN is verified by extensive experiments under the premise of whether the skeleton has physical relationship constraints or not.Compared with the existing advanced algorithms,the accuracy of action recognition of AS-GCN is improved by 12% and 4.84% respectively in CV and CS angles on average with skeleton physical constraints.While under the condition of no skeleton physical constraint,the average improvement is 20% and 14.49% in CV and CS angles,respectively.

Progressive Multi-stage Image Denoising Algorithm Combining Convolutional Neural Network and
Multi-layer Perceptron

XUE Jinqiang, WU Qin

Computer Science. 2024, 51 (4): 243-253. doi:10.11896/jsjkx.230100140

Abstract

PDF(3687KB) ( 97 )

References | Related Articles | Metrics

Among the existing image denoising methods based on deep learning,there are problems at the network architecture dimension that single-stage network is hard to represents feature dependency and it is difficult to reconstruct clear images in complex scenarios.The internal features of multi-stage networks are not tightly connected and the original image details are easily lost.At the basic building block dimension,there are problems that the convolutional layer is difficult to handle cross-level features at large noise levels,and the fully connected layer is difficult to capture the spatial details of the image locality.To solve the above problems,this paper proposes solutions from two aspects.On the one hand,a novel cross-stage gating feature fusion is proposed at the architecture dimension,so as to better connect the shallow features of the first-stage network with the deep features of the second-stage network,promote the interaction of information flow and make the internal correlation of the denoising network closer,while avoiding the loss of original spatial details.On the other hand,a dual-axis shifted block combining convolu-tional neural network(CNN) and multi-layer perceptron(MLP) is proposed,which is applied to low-resolution and multi-channel number feature maps to alleviate the problem of insufficient learning ability of CNN on cross-level feature dependencies in complex noise scenarios.And CNN is used to focus on high-resolution feature maps with low channel number to fully extract the spatial local dependencies of noisy images.Many quantitative and qualitative experiments prove that the proposed algorithm achieves the best peak signal-to-noise ratio(PSNR) and structural similarity(SSIM)denoising indicators with a small number of parameters and computational costs in real-world image denoising and Gaussian noise removal tasks.

Global Covariance Pooling Based on Fast Maximum Singular Value Power Normalization

ZENG Ruiren, XIE Jiangtao, LI Peihua

Computer Science. 2024, 51 (4): 254-261. doi:10.11896/jsjkx.230200140

Abstract

PDF(1950KB) ( 56 )

References | Related Articles | Metrics

Recent research work shows that matrix normalization plays a key role in global covariance pooling,which helps to generate more discriminative representations,thus improving the performance of image recognition tasks.For different normalization methods,the matrix structure-wise normalization can make full use of the geometric structure of the covariance matrix,so it can obtain better performance.However,the structure-wise normalization generally depends on singular value decomposition(SVD) or eigenvalue decomposition(EIG) with high computational cost,which limits parallel computing ability of GPUs,beco-ming a computational bottleneck.Iterative matrix square root normalization(iSQRT) uses Newton-Schulz iteration to normalize the covariance matrix,which is faster than the methods based on SVD and EIG.However,with the increase of the number of itera-tions and dimensions,the time and memory of iSQRT will increase significantly,and this method cannot complete the normalization of general power,which limits its application scope.To solve the above problems,a covariance matrix normalization method based on the maximum singular value power is proposed by dividing the covariance matrix by the power of its maximum singular value which only depends on iterative power method to estimate the maximum singular value of the matrix.Detailed ablation experiments show that,compared with iSQRT,the proposed method is faster and occupies less memory,and is superior to iSQRT in terms of time complexity and space complexity,and its performance is comparable to or better than iSQRT.The proposed method has achieved state-of-the-art performance in large-scale image classification dataset and fine-grained visual recognition datasets,including Aircraft,Cars and Indoor67,where accuracy is 90.7%,93.3% and 83.9% respectively.The result fully demonstrates the robustness and generalization of the proposed method.

Speech Emotion Recognition Based on Voice Rhythm Differences

ZHANG Jiahao, ZHANG Zhaohui, YAN Qi, WANG Pengwei

Computer Science. 2024, 51 (4): 262-269. doi:10.11896/jsjkx.230200063

Abstract

PDF(2489KB) ( 71 )

References | Related Articles | Metrics

Speech emotion recognition has an important application prospect in financial anti-fraud and other fields,but it is increasingly difficult to improve the accuracy of speech emotion recognition.The existing methods of speech emotion recognition based on spectrograms are difficult to capture the rhythm difference features,which affects the recognition effect.Based on the difference of speech rhythm features,this paper proposes a speech emotion recognition method based on energy frames and time-frequency fusion.The key is to screen high-energy regions of the spectrum in the speech,and reflect the individual voice rhythm differences with the distribution of high-energy speech frames and time-frequency changes.On this basis,an emotion recognition model based on convolutional neural network(CNN) and recurrent neural network(RNN) is established to realize the extraction and fusion of the time and frequency changes of the spectrum.On the open data set IEMOCAP,the experiment shows that compared with the method based on spectrogram,the weighted accuracy WA and the unweighted accuracy UA of the speech emotion recognition based on the difference of speech rhythm increases by 1.05% and 1.9% on average respectively.At the same time,it also shows that individual voice rhythm difference plays an important role in improving the effect of speech emotion recognition.

Survey of UAV-assisted Energy-Efficient Edge Federated Learning

LU Yanfeng, WU Tao, LIU Chunsheng, YAN Kang, QU Yuben

Computer Science. 2024, 51 (4): 270-279. doi:10.11896/jsjkx.231100084

Abstract

PDF(2006KB) ( 96 )

References | Related Articles | Metrics

With the rapid development of mobile communication technology and the proliferation of Internet of Things(IoT) terminal devices,rich and diverse intelligent applications and massive data are generated at the edge of the network,and edge intelligence applications are born.Currently,as an emerging distributed machine learning method,federated learning can collaborate to complete the model training task without sharing the raw data of terminal devices,which is an important way to achieve edge intelligence.The traditional edge intelligence network uses the ground communication base station as the parameter server,and its service range is relatively fixed,which cannot adapt to the complex and changing heterogeneous network environment.Unmanned aerial vehicles(UAVs) introduced into federated learning due to their flexibility and mobility,so as to effectively provide communication/computation/caching services in edge intelligence networks,enhance the communication capacity of the ground network,and make up for the shortcomings of the traditional ground network such as limited communication range,high communication overhead,and high data transmission delay.UAV-assisted federated learning has obvious advantages such as wide communication coverage,low communication overhead,and instant response,but it also faces challenges such as limited communication bandwidth,unreliable communication environment,and uncertainty of flight environment,and the above challenges may lead to low energy efficiency problems.UAV-assisted energy efficient edge federated learning is to study the energy efficiency optimization scheme by considering the computational energy consumption,computational frequency and time allocation of UAVs as edge ser-vers.For the scenario of UAVs as edge servers,the current research on UAV-assisted energy-efficient federated learning is classified and summarized on the basis of different optimization objectives,such as minimizing energy consumption,minimizing latency,and minimizing energy-delay weighted sums,and the future research directions are considered and outlooked.

Multi-agent Reinforcement Learning Method Based on Observation Reconstruction

SHI Dianxi, HU Haomeng, SONG Linna, YANG Huanhuan, OUYANG Qianying, TAN Jiefu , CHEN Ying

Computer Science. 2024, 51 (4): 280-290. doi:10.11896/jsjkx.230600055

Abstract

PDF(3228KB) ( 84 )

References | Related Articles | Metrics

Common knowledge is a well-known knowledge set within a multi-agent system.How to make full use of common knowledge for strategic learning is a challenging problem in multi-agent independent learning systems.In addressing this pro-blem,this paper proposes a multi-agent reinforcement learning method called IPPO-CKOR based on observation reconstruction,focusing on common knowledge extraction and independent learning network design.Firstly,the common knowledge features of agents' observation information are computed and fused to obtain fused observation information with common knowledge features.Secondly,an agent selection algorithm based on common knowledge is used to select closely related agents,and a feature generation mechanism based on reconstruction is employed to construct their feature information.The reconstructed observation information,composed of the fused observation information with common knowledge features,is utilized for learning and executing agent policies.Thirdly,a network structure based on observation reconstruction is designed,which employs multi-head self-attention mechanism to process the reconstructed observation information and uses one-dimensional convolution and GRU layers to handle observation information sequences.This enables the agents to extract more effective features from the observation information sequences,effectively alleviating the impact of non-stationary environments and partially observable problems.Experimental results demonstrate that the proposed method outperforms existing typical multi-agent reinforcement learning methods that employ independent learning in terms of performance.

Study on Open Set Activity Recognition Technology Based on Wearable Devices

WANG Jiahao, YAN Hang, HU Xin, ZHAO Dexin

Computer Science. 2024, 51 (4): 291-298. doi:10.11896/jsjkx.230300158

Abstract

PDF(3486KB) ( 61 )

References | Related Articles | Metrics

With the popularity of wearable devices such as smart watches and bracelets,using them for human activity recognition and decoding human behavior is of great significance for health monitoring,daily behavior analysis,smart home and other applications.However,traditional action recognition algorithms have problems such as difficult feature extraction and low recognition accuracy,and are all based on the close set assumption,that is,all training data and test data come from the same label space,while most of the real world is open.In the open-set scene,unknown label samples may be sent to the model during the test phase,resulting in incorrect classification.This paper proposes a multi-channel adaptive convolutional network(MCACN) for human acti-vity recognition.For the problem that the traditional CNN network feature extraction is limited to a small range,the adaptive convolution module can use convolution kernels of different sizes to extract features of different time spans,automatically calculate the weights and sum them up.In addition,the multi-channel structure of MCACN enables each sensor data to be processed separately to obtain feature details that can distinguish similar actions.Finally,this paper designs a label-based multivariate variational autoencoder,and proposes MCACN-VAE for open set recognition.The model can identify unknown classes by calculating recons-truction loss,focusing on known class actions,and improving the robustness of the model.Experimental results show that in the closed set experiment,the MCACN model can effectively recognize the actions,and the accuracy of the recognition of seven daily actions has reached more than 91%,the overall accuracy has reached 95%.In the open set experiment,the overall recognition accuracy of MCACN-VAE for known categories has reached more than 89% at different degrees of openness,and the recognition accuracy of unknown action segments has also remained above 75%.It proves that the proposed model can effectively reject unknown classes and identify known classes.

Unified Fake News Detection Based on Semantic Expansion and HDGCN

ZHANG Mingdao, ZHOU Xin, WU Xiaohong, QING Linbo, HE Xiaohai

Computer Science. 2024, 51 (4): 299-306. doi:10.11896/jsjkx.230700170

Abstract

PDF(1627KB) ( 58 )

References | Related Articles | Metrics

here are many methods for detecting fake news.The single method typically focuses only on information such as news content,social context,or external facts.On the other hand,joint detection methods integrate multiple modalities of information to achieve the detection goal.Pref-FEND is an example of a joint detection method that integrates news content and external facts.It extracts three types of word representations from news content and external facts,and uses dynamic graph convolutional networks to capture relationships between word nodes.However,there are still shortcomings in how to make each modality more focused on its preferred aspect.Therefore,the Pref-FEND model has been improved by using semantic mining to expand style words in news and entity linking to expand entity words in news.This results in five types of word as node representations in the graph neural network,enabling a more effective modeling of the node representation of the graph neural network.Additionally,a deep heterogeneous graph convolutional network(HDGCN) is introduced for preference learning.Its deep strategy and multi-layer attention mechanism allow both models to focus more on their own preferred perception and reduce redundant information.Experimental results demonstrate the effectiveness of the improved framework.On the public datasets Weibo and Twitter,compared to the current state-of-the-art content-based single model LDAVAE,the improved framework achieves an F1 score improvement of 2.8% and 1.9% respectively.Compared to the fact-based single model GET,the F1 score improvement is 2.1% and 1.8% respectively.In the case of joint detection using LDAVAE+GET,the F1 score is 1.1% and 1.3% higher than Pref-FEND respectively.Experimental results validate the effectiveness of the improved model.

Study on Improved Fake Information Detection Method Based on Cross-modal CorrelationAmbiguity Learning

DUAN Yuxiao, HU Yanli, GUO Hao, TAN Zhen, XIAO Weidong

Computer Science. 2024, 51 (4): 307-313. doi:10.11896/jsjkx.230900087

Abstract

PDF(1990KB) ( 53 )

References | Related Articles | Metrics

In recent years,with the rapid development of the Internet and multimedia technology,it is more convenient for people to obtain information,but the spread of fake information on the Internet is also increasingly serious,and the negative impact is constantly expanding.In order to enhance the credibility and deception,fake information presents a multi-modal development trend,which makes the detection work face greater challenges.The existing multi-modal fake information detection methods pay more attention to the formation of multi-modal features.The research on the contribution rate of cross-modal ambiguity and different modal features in detection is not perfect,ignoring the impact of inherent differences among different modal features on fake information detection.To solve the problem,this paper proposes to construct an improved fake information detection model based on cross-modal correlation ambiguity learning.Through cross-modal ambiguity learning of text and image features,the weights of unimodal features and fused features are updated by the ambiguity score.The unimodal features and fused features are combined adaptively,and the weights of text and image features are dynamically assigned by grid search to improve the detection accuracy.The effectiveness of the model is verified by experiments on the Twitter dataset.The accuracy is improved by 6% compared with the baseline model and 1.6% compared with the detection without dynamic weight assignment.

Fake Review Detection Based on Residual Networks Fusion of Multi-relationship Review Features

LUO Zeyang, TIAN Hua, DOU Yingtong, LI Manwen, ZHANG Zehua

Computer Science. 2024, 51 (4): 314-323. doi:10.11896/jsjkx.230200020

Abstract

PDF(3973KB) ( 70 )

References | Related Articles | Metrics

With the rise of e-commerce and short video community platforms,the emergence of fake reviews has seriously affected user experience.Even to combat platform detection,review camouflage makes it harder to distinguish between true and false.Current fake review detection methods based on graph neural networks(GNNs) are prone to network degradation and gradient disappearance during deep training.At the same time,review camouflage causes review markers to skew more,which affects the robustness of GNNs detection model.To solve the above problems,a detection method based on residual network(MRDRN) is proposed,which can fuse the features of multi-relationship reviews to identify fake reviews.Firstly,in order to slow down network degradation,the feature extraction of deep reviews is carried out by combining residual network.A new neighbor mixed sampling strategy is proposed,which can be used to conduct low-and high-order neighbor mixed sampling according to the feature similarity between reviews,so as to alleviate the problem of imbalanced review marks and learn more rich review features.Secondly,a multi-relationship review features fusion strategy is proposed,which reduces the impact of review masking by integrating intra relationship review network topology and inter relationship review features as a whole.Experimental results on three real datasets show that MRDRN has higher detection capability and stronger robustness than the standard method.

Multilingual Opinion Factor Extraction Fusing Aspect Semantics and Grid Tagging

GU Wenxia, ZAOKERE Kadeer, YANG Qian, AISHAN Wumaier

Computer Science. 2024, 51 (4): 324-333. doi:10.11896/jsjkx.230200195

Abstract

PDF(2355KB) ( 48 )

References | Related Articles | Metrics

Aspect-oriented fine-grained opinion extraction(AFOE) aims to extract the aspect and opinion terms in the reviews in the form of opinion pairs or to extract the sentiment polarity to form opinion triplets.Previous studies usually extract opinion factors in a pipeline manner,which is prone to the problem of error propagation,most of them only focus on the single sub-task extraction of aspect terms or opinion terms,and ignore the mutually interactive and indicative information between different opinion factors,which lead to the problem that opinion excavation tasks are incomplete.In addition,the existing researches do not pay attention to the research of Chinese-oriented opinion factors extraction.To tackle the problems,this paper proposes multilingual opinion factors extraction model fusing aspect semantics and grid tagging.Firstly,inward LSTM(Inward-LSTM) and outward LSTM(Outward-LSTM) are exploited to encode aspect terms and corresponding left-right contexts to establish the association between aspect and candidate opinion terms,and then combine global context information to generate contextualized representation of specific aspect semantic features,which is beneficial to improve the performance of downstream opinion factors extraction.Secondly,the inference strategy of the grid tagging scheme is applied to decode the potential indications between aspect and opi-nion terms for more accurate extraction,the AFOE task is handled in an end-to-end manner.Compared with the baseline model,the F1 scores of the proposed model in the Chinese and English datasets increase by 0.89%~4.11% for the aspect opinion pair extraction task,and 1.36%~3.11% for the triplet extraction task.Experimental results show that the improved model can effectively extract the opinion factors of Chinese and English comments,the performance is significantly better than the baseline model.

Study on Unmanned Vehicle Trajectory Planning in Unstructured Scenarios

ZHU Wei, YANG Shibo, TENG Fan, HE Defeng

Computer Science. 2024, 51 (4): 334-343. doi:10.11896/jsjkx.221200079

Abstract

PDF(5645KB) ( 61 )

References | Related Articles | Metrics

Aiming at the problems of low real-time performance and poor track smoothness of traditional unmanned vehicle tra-jectory planning algorithm in unstructured scenes,this paper proposes a front and rear separated trajectory planning algorithm.The front-end path search part of the algorithm prunes the search range of the Hybrid A^* algorithm in the control space and retains the kinematic constraints of the vehicle,and improves the real-time performance of the graph search by optimizing the calculation method of the heuristic function.The back-end trajectory optimization part of the algorithm is divided into two stages:in the first stage,a soft-constrained nonlinear multi-objective optimizer is designed to locally optimize the path and generate discrete trajectory pose points and time allocation values;in the second stage,based on the quintic spline uses the idea of minimizing Jerk to smoothly connect the discrete pose points,which improves the smoothness of the trajectory.Finally,the proposed algorithm is tested on a real vehicle in an outdoor parking lot environment.Experimental results of front-end path search and back-end trajectory optimization show that the algorithm has high real-time performance and trajectory smoothness.

Deep Reinforcement Learning Portfolio Model Based on Dynamic Selectors

ZHAO Miao, XIE Liang, LIN Wenjing, XU Haijiao

Computer Science. 2024, 51 (4): 344-352. doi:10.11896/jsjkx.230100048

Abstract

PDF(3132KB) ( 99 )

References | Related Articles | Metrics

In recent years,portfolio management problems have been extensively studied in the field of artificial intelligence,but there are some improvements in the existing quantitative trading methods based on deep learning.First of all,the prediction model of stocks is single,usually a model only trains a trading expert,and the decision of trading is only based on the prediction results of the model.Secondly,the data source used in the model is relatively single,only considering the stock's own data,ignoring the impact of the entire market risk on the stock.Aiming at the above problems,a reinforcement learning model based on dynamic selection predictor(DSDRL) is proposed.The model is divided into three parts.Firstly,the characteristics of stock data are extracted and introduced into multiple predictors.Multiple prediction models are trained for different investment strategies,and the current optimal prediction results are obtained by dynamic selector.Secondly,the market environment evaluation module is used to quantify the current market risk and obtain the appropriate proportion of investment amount.Finally,based on the first two mo-dules,a deep reinforcement learning model is established to simulate the real trading environment,and the actual portfolio strategy is obtained based on the predicted results and the proportion of investment amount.In this paper,the daily k-line data of China Securities 500 and S & P 500 are used for test verification.The results show that the proposed model is superior to other refe-rence models in Sharpe rate and other indicators.

Study on Automatic Classification of English Tense Exercises for Intelligent Online Teaching

TU Xin, ZHANG Wei, LI Jidong, LI Meijiao , LONG Xiangbo

Computer Science. 2024, 51 (4): 353-358. doi:10.11896/jsjkx.240300109

Abstract

PDF(2051KB) ( 53 )

References | Related Articles | Metrics

With online teaching becoming one of the normalized teaching methods,people put forward higher quality teaching demands.Various online teaching platforms and the amount of educational resources on the Internet have greatly facilitated many learners.However,there are also some problems in educational resources such as uneven quality,lack of effective classification and integration,and mainly rely on manual sorting,which lead to people spending too much time and energy to search,screen and sort online educational resources.Considering the existing shortcomings of online education resources,this paper proposes an automatic classification method for online education resources based on natural language processing technology,and conduct experiments on the automated classification of eight English tense exercises,which are the key contents of middle school English grammar teaching.The experiment collects more than 90 000 English tense exercises both online and offline.After data cleaning,approximately 30 000 sentences are selected to construct a dataset,and a BERT fine-tuning text classification model is constructed.By training the model,automatic classification of the eight tenses is realized with an overall classification accuracy of 86.15%.And the recognition accuracy for the present tense is the highest,reaching 93.88%.To a certain extent,in terms of English tenses,the experimental result can meet the practical needs of automatic classification and organization of English education resources,intelligent correction and personalized push of exercises,intelligent Q&A.It provides a feasible idea and solution for improving the quality of online teaching and integrating online education resources.

Multi-generator Active Learning Algorithm Based on Reverse Label Propagation and ItsApplication in Outlier Detection

XING Kaiyan, CHEN Wen

Computer Science. 2024, 51 (4): 359-365. doi:10.11896/jsjkx.230500034

Abstract

PDF(2090KB) ( 52 )

References | Related Articles | Metrics

The current problem of unbalanced distribution of positive and negative training samples has greatly limited the performance of outlier detection models.The outlier detection algorithm based on active learning can automatically synthesize outliers to balance the training data through active learning of sample distribution.However,the traditional detection method based on active learning lacks the quality assessment and filtering of synthetic outliers,which leads to the fact that the noise in the synthetic training samples degrades the performance of classification models.Aiming at the above problems,a multi-generator adversarial learning algorithm based on reverse label propagation(MG-RLP) is proposed,which consists of multiple neural network generators and a discriminator for outlier boundary detection.MG-RLP uses multiple sub-generators to generate sample data with multi-distribution features to prevent the mode collapse problem caused by the excessive aggregation of training samples synthesized by a single generator.At the same time,the proposed method utilizes the reverse label propagation to evaluate the quality of the sample points generated to screen out credible synthetic samples.The filtered samples are retained in the training samples to iteratively train the discriminator to improve the detection performance of outliers.The MG-RLP is compared with six typical outlier detection algorithms on five public datasets.The results show that the proposed algorithm improves AUC and detection precision by 15% and 22% respectively,which verifies its effectiveness.

Security Scheme of UAV Flight Control Based on Attribute Access Control Policy

PANG Yuxiang, CHEN Zemao

Computer Science. 2024, 51 (4): 366-372. doi:10.11896/jsjkx.230200135

Abstract

PDF(2079KB) ( 67 )

References | Related Articles | Metrics

The flight control system is the core component of unmanned aerial vehicles(UAVs),which plays a decisive role in the function and performance,and it is a crucial target for information security protection.In this paper,a location-and-environment oriented attribute-based access control(LE-ABAC) policy is designed to deal with the security risks of malicious code injection and internal interactive data tampering faced by PX4 flight control system.The access control policy,based on object entity attri-butes and external location environment information of the UAV,formulates corresponding rules that enable fine-grained control of the data exchange process within the UAV,protecting the confidentiality and integrity of crucial data exchanges.In the study,attack simulation experiments are conducted on the PX4 software simulation platform to verify the proposed scheme.Finally,the results show that the model can effectively protect the interactive data of the flight control system from theft and tampering without significantly reducing the efficiency of UAV flight control execution.

Active Membership Inference Attack Method Based on Multiple Redundant Neurons

WANG Degang, SUN Yi, GAO Qi

Computer Science. 2024, 51 (4): 373-380. doi:10.11896/jsjkx.230100024

Abstract

PDF(3262KB) ( 51 )

References | Related Articles | Metrics

Federated learning provides privacy protection for source data by exchanging model parameters or gradients.However,it still faces the problem of privacy disclosure.For example,membership inference attack can infer whether the target data samples are used to train machine learning models in federated learning.Aiming at the problem that the existing active membership inference attack based on model parameter construction in federated learning are less robust to dropout operations,an active membership inference attack method is proposed.This method makes use of the characteristic that the input of ReLU activation function is negative and the output is zero,constructs model parameters according to the target data,and inferences membership through the difference between member data and non-member data in updating model parameters.The redundancy of model neurons is used to construct multiple paths to achieve robustness to dropout.Experiments on MNIST,CIFAR10 and CIFAR100 datasets proves the effectiveness of our method.When dropout is used in model training,the proposed method can still achieve an accuracy of 100%.

Study on Trust Management Mechanism of Internet of Vehicles Based on Blockchain

LI Fengyun, CHEN Mingming, WANG Lin, LI Peng , JU Xianyin

Computer Science. 2024, 51 (4): 381-387. doi:10.11896/jsjkx.230900057

Abstract

PDF(2617KB) ( 120 )

References | Related Articles | Metrics

With the development of autonomous driving and intelligent transportation systems,vehicle networking technology is playing an increasingly important role.Due to the open access environment of the Internet of Vehicles,how to ensure the reliability of messages and the credibility of vehicles has become a major security challenge.Building upon existing blockchain-based trust management solutions,there is a need to redesign a trust management framework for vehicular networks to address scalability issues and the inefficiency of consensus algorithms in current solutions.The framework is primarily composed of three modules:message trust evaluation,vehicle trust update,and the creation and consensus of trust blocks.In the message trust evaluation module,to identify false messages from malicious nodes,the credibility of messages is comprehensively assessed based on the direct trust of vehicle entities and the indirect trust from neighboring vehicles.In the vehicle trust update module,to effectively curb malicious behavior,vehicle trust is adjusted based on message evaluation results and the historical behavior of vehicles.In the block creation and consensus module,an optimized consensus algorithm based on proof of importance is proposed,considering event significance and blockchain scalability.Finally,the usability of the framework is verified by simulation experiments,and the comparative experimental results show that the proposed algorithm achieves good results in scalability and robustness.

Android Malware Detection Method Based on GCN and BiLSTM

HE Jiaojun, CAI Manchun, LU Tianliang

Computer Science. 2024, 51 (4): 388-395. doi:10.11896/jsjkx.230100002

Abstract

PDF(2729KB) ( 89 )

References | Related Articles | Metrics

Most of the existing Android malware detection methods learn features of a single structure type,and there are shortcomings in analyzing application semantics.Aiming at the problem that the traditional detection methods are not comprehensive enough in capturing feature semantics,this paper innovatively proposes an Android malware detection model based on GCN and BiLSTM.At the same time,the semantic of malicious behavior is analyzed emphatically while the sample structure information is extracted accurately.Firstly,the topological relationship between 26 types of key system calls is represented in the graph,and the two-layer GCN network is used to aggregate the high-order structure information of nodes in the system call graph to effectively improve the feature learning efficiency.Then,the BiLSTM network with self-attention mechanism is used to obtain the context semantics of opcode sequence.By assigning high weights to sequences with malicious features,the strong correlation within features is obtained.Finally,Softmax is used to output the sample classification probability fused with structural information and context features.In the experiments based on Drebin and AndroZoo datasets,the accuracy of the proposed model reaches 93.95%,and the F1 value reaches 0.97,which is significantly improved compared with the benchmark algorithm.It fully proves that the proposed model based on GCN and BiLSTM can effectively discriminate the properties of applications and improve the detection effect of Android malware.