Computer Science

CONTENTS

Computer Science. 2024, 51 (10): 0-0.

Abstract

PDF(430KB) ( 405 )

RelatedCitation | Metrics

Metaverse Teaching:A Higher Form of Digital Teaching Transformation in Higher Education

ZHANG Ce, CHU Dianhui, ZHANG Qiao, LIU Peng, WEI Meng, LIU Xiaoying

Computer Science. 2024, 51 (10): 1-9. doi:10.11896/jsjkx.240400083

Abstract

PDF(2968KB) ( 468 )

References | Related Articles | Metrics

With the new technological revolution,information technology is rapidly changing human thinking and actions,promoting changes in higher education and teaching in the digital era.The digital transformation of higher education has become the greatest driving force for the adaptation and leading development of higher education and teaching.In this wave of digital transformation in education and teaching,the metaverse education and teaching has received widespread attention due to its imaginative features such as virtual agility,wonderful vividness,and intelligent linkage.The origin of the metaverse is outlined and the educational metaverse is analyzed.Focusing on the basic characteristics,theoretical basis,mapping and interaction between “virtual” and “real”,learning methods,as well as basic classification and process of situational and scenario based teaching,this paper analyses the teaching of the metaverse,discusses the teaching cases and practical progress of the metaverse.The preliminary architecture and core functions of the metaverse teaching platform is provided,and a thorough systematic explanation is given.Finally,it analyzes the intelligent meta-universe teaching enabled by AI technology,and points out that it opens up new avenues for curriculum construction and teaching reform.In terms of teaching resources and content,environment and space,organization and form,methods and technologies,new forms will be given,providing new choices for teaching and learning.It explaines the basic logic of promoting the digital transformation of higher education through intelligent metaverse teaching,and analyzes the technologies that urgently need to be broken through to restrict the widespread application of metaverse teaching.The promising future of teaching in the metaverse urgently requires more in-depth research in concepts,ideas,and theories,as well as synchronous or even advanced exploration and experimentation in practice.It requires the academic community of education and teaching,the scientific and technological innovation technology community,and the research and development manufacturing industry to work together to create a new future for the transformation of digital teaching in universities.

Computational Perception Technologies in Intelligent Education:Systematic Review

LIU Feng, LIU Yaxuan, CHAI Xinyu, JI Haohan, ZHENG Zhixing

Computer Science. 2024, 51 (10): 10-16. doi:10.11896/jsjkx.240400112

Abstract

PDF(1929KB) ( 327 )

References | Related Articles | Metrics

With the advancement and development of educational technology,computational perception technologies such as deep learning,machine learning,virtual reality,and the Internet of Things are increasingly being applied in human-centric,empathetic education,playing a key role in empowering educational transformation.However,there is still limited knowledge regarding the application of these technologies in intelligent education.This study systematically analyzes the progress and applications of computational perception technologies in intelligent education,drawing on perceptible data from both physical and virtual spaces,such as facial expressions,speech,text,eye movements,touch,and physiological signals.This study conducts a statistical analysis and screening of currently published journal articles and conference papers.The study explores the advancements and applications of these technologies,discussing their potential impacts and challenges in educational practice.Finally,a forward-looking discussion on the future development directions of computational perception technologies in intelligent education is presented based on the conclusions of this study.

Survey on Deep Learning-based Personalized Learning Resource Recommendation

ZHOU Yangtao, CHU Hua, ZHU Feifei, LI Xiangming, HAN Zihan, ZHANG Shuai

Computer Science. 2024, 51 (10): 17-32. doi:10.11896/jsjkx.240400088

Abstract

PDF(2115KB) ( 485 )

References | Related Articles | Metrics

With the deep integration of information technology and education,novel online education,as a pivotal component of smart education,provides learners with convenient online e-learning platforms and rich learning resources.However,the rapid development of online education modes has also led to significant challenges such as “knowledge overload” and “knowledge dis-orientation”,which severely limits learners' educational gains and efficiency.In recent years,learning resource recommendation,as a key technology for information filtering,aims to analyze learners' historical behaviors,capture their underlying learning needs,and ultimately achieve personalized learning resource recommendation services.Accurate personalized learning resource recommendations can effectively address the challenges of “knowledge overload” and “knowledge disorientation” in online education,making it an indispensable core function in major online e-learning platforms.In addition,with the continuous advancement of deep learning technologies,research on deep learning-based personalized learning resource recommendation has become a focal area of interdisciplinary study in computer science and education.Therefore,this paper systematically analyzes existing research from multiple dimensions and levels,guided by the research questions of “how to achieve personalized learning resource recommendations” and “how to evaluate recommendation results”.Specifically,the paper firstly categorizes and summarizes the per-sonalized recommendation process of learning resources from five dimensions,including characteristics,recommendation objectives,deep learning technologies,integration methods of side information,and recommendation patterns,to answer the question of how to realize personalized recommendation of learning resources.Second,this paper inductively compares the evaluation process of recommendation results from three aspects,including datasets,evaluation metrics,and experimental methods,and provides unified download links for all open-source datasets,to answer the question of how to evaluate the recommendation results.Finally,this paper explores future research trends of learning resource recommendation from two perspectives:overcoming the inherent limitations of current recommendation methods as well as integrating and utilizing external emerging technologies.

Survey of Research on Automated Grading Algorithms for Subjective Questions

FENG Jun, LI Kaixuan, GAO Zhizezhang, HUANG Li, SUN Xia

Computer Science. 2024, 51 (10): 33-39. doi:10.11896/jsjkx.240400008

Abstract

PDF(1592KB) ( 354 )

References | Related Articles | Metrics

In educational teaching,paper assessment is an important means for teachers to understand students' grasp of know-ledge points.However,grading exam questions is a time-consuming process,and assessing subjective questions requires examiners to review the work carefully,with engagement and attention to detail,often consuming a significant amount of energy.To reduce the workload on teachers and improve the efficiency of subjective question assessment,research on AI-based automatic grading techniques is imperative,with subjective question evaluation posing a particular challenge.With advancements in machine learning and deep learning in the field of natural language processing,significant progress has been made in the automation of subjective question assessment.This paper categorizes subjective questions into conventional and open-ended types,respectively,conducts a literature review,summarizes evaluation criteria and publicly available datasets,and outlines methods and technological approaches involved.Finally,the future research directions of automatic evaluation of subjective questions is summarized and prospected.

Survey on Intelligent Analysis Techniques for Classroom Teacher-Student Interaction Research

CUI Jiajun, KANG Lu, MA Miao

Computer Science. 2024, 51 (10): 40-49. doi:10.11896/jsjkx.240400084

Abstract

PDF(1739KB) ( 315 )

References | Related Articles | Metrics

With the popularization and continuous development of education informatization,a huge amount of classroom data such as video,image,voice,text are recorded.How to effectively analyze these multimodal data and mine the classroom teacher-student interaction information can not only help teachers find the problems in teaching and adjust the teaching content in time to improve the quality of teaching,moreover,it is an important link to implement the concept of “human-centered” education and promote modern education towards intelligence,personalization and digitalization.The paper firstly discusses the traditional ana-lysis methods of teacher-student interaction behaviors at home and abroad.Then,it classifies and analyzes the current research status of intelligent analysis techniques for classroom teacher-student interaction from different perspectives,such as video,image,voice,text and multimodal.Next,a technical process for classroom teacher-student interaction intelligent analysis is proposed,including core elements,data forms,key technologies,results presentation and application scenarios.Finally,the advantages and disadvantages of the current multimodal intelligent analysis technology for classroom teacher-student interaction are summarized,as well as the challenges and future directions.

Crowdsourced Learning Method and Its Applications in Course Teaching

MAO Xinjun, LU Yao

Computer Science. 2024, 51 (10): 50-55. doi:10.11896/jsjkx.240300033

Abstract

PDF(2090KB) ( 270 )

References | Related Articles | Metrics

Current education and teaching are mostly closed in physical boundaries,highly depending on teachers and textbooks,and deficient in the reuse of learning knowledges.For these limitations,we propose the concept and method of crowdsourced learning(CL) by drawing on the ideas of crowd intelligence and open source software.The core idea of CL is to break the boundaries of traditional classes,grades,and schools,and allow learners to engage in autonomous and collaborative social learning around specific themes in the form of open communities,such as problem discussions,experience sharing,and resource co-production.Such methods can produce a large number of high-quality and personalized learning resources,help learners effectively conduct course learning and practice,while also help teachers improve the efficiency and quality of course teaching.This paper also introduces the platform called LearnerHub that supports the CL,analyzes the application patterns of CL based on the software engineering course practices.We also evaluate the effectiveness and influence of CL in term of data analyses and investigation survey.The results show that students highly recognize the important role of CL method in course studies and practices,and there is a positive relationship between students' comprehensive practice performance of CL and their course grades.

Perception and Analysis of Teaching Process Based on Video Understanding

DUAN Xinran, WANG Mei, HAN Tianli, ZHOU Hongyu, GUO Junqi, JI Weixing, HUANG Hua

Computer Science. 2024, 51 (10): 56-66. doi:10.11896/jsjkx.240400109

Abstract

PDF(4323KB) ( 329 )

References | Related Articles | Metrics

The classroom serves as the core battleground for education.Monitoring and evaluating teachers' instructional activities in the classroom is an effective means of improving the quality of teaching.However,existing manual evaluation methods suffer from drawbacks such as low efficiency,potential disruption of classroom dynamics,and subjective errors,making it difficult to achieve satisfactory results.Given the rapid development of artificial intelligence(AI) technology,it is proposed to integrate human-centered intelligent analysis techniques into teachers' instructional processes for real-time recognition and analysis of tea-chers.First,a facial detection algorithm is employed to locate the teacher's position and estimate their movements.Second,a gaze estimation algorithm is utilized to detect the teachers' focal points.Lastly,skeleton-based action recognition and facial expression recognition are employed to perceive and analyze teachers' actions and expressions.Quantitative statistics on the indicators provide a more efficient and objective understanding of teachers' teaching characteristics,so as to help teachers improve their tea-ching quality.As experimented in the same configuration environment,the modules of the system perform well in the correspon-ding tasks and fulfill the requirements in teaching scenarios.From the evaluation results on real-world teaching videos,the system is designed to accurately perceive the teachers' instructional states,providing constructive feedback for enhancing teaching quality.

Learning Pattern Recognition and Performance Prediction Method Based on Learners'Behavior Evolution

HUANG Chunli, LIU Guimei, JIANG Wenjun, LI Kenli, ZHANG Ji, TAK-SHING Peter Yum

Computer Science. 2024, 51 (10): 67-78. doi:10.11896/jsjkx.240500002

Abstract

PDF(3403KB) ( 289 )

References | Related Articles | Metrics

Online learning provides learners with open and flexible learning opportunities,but suffers low learning engagement and unsatisfactory academic performance.Existing works on academic performance prediction mainly study how behaviors will impact performance from a static perspective,and neglect learners' behavior evolution over time and lack a deep understanding of learning patterns and learners' motivations,which are the key factors in learning performance.Therefore,a method of perfor-mance prediction based on learners' learning pattern and motivation is proposed to model the effects of learners' patterns and motivations on their performances.First,we quantify learning efficiency based on learners' efforts and gains and construct the dynamic evolution sequence of learning efficiency with time.Then,we cluster learners' behavior and identify four typical learning patterns combined with the actual learning scenarios.Based on this,learning pattern recognition and motivation prediction mode-ling are designed.The final performance prediction model is constructed by combining them with the bi-directional long-and short-term memory networks.Furthermore,we conduct a detailed and in-depth data analysis on each type of learning patterŃs efforts and gains in eight online courses.Comparative experiments show that the proposed model performs better on several metrics,with improvements ranging from 6.9% to 29.2%.Our work will help online learners,teachers,and platforms accurately understand learners' learning states and improve online learning performance.

Key Information Retrieval System for MOOC Videos

ZHAO Bocheng, BAO Lantian, YANG Zhesen, CAO Xuan, MIAO Qiguang

Computer Science. 2024, 51 (10): 79-85. doi:10.11896/jsjkx.240400087

Abstract

PDF(2590KB) ( 268 )

References | Related Articles | Metrics

Thanks to the rapid advancement of Internet technology,online education platforms,particularly massive open online courses(MOOCs),have increasingly captured public attention.MOOCs represent a revolutionary educational approach,effectively eliminating the geographical boundaries inherent in traditional education models and fostering the worldwide dissemination of elite educational resources.These courses empower learners to cherry-pick courses based on their unique interests,create flexible study schedules,monitor their progress,and revisit materials as needed.Despite their versatility,current MOOC platforms still struggle to pinpoint precise knowledge nuggets within lecture videos.This often leads learners to constantly scrub through the video timeline,searching for relevant segments,thereby disrupting the learning continuum.In view of this situation,we introduce a MOOC video knowledge extraction algorithm,leveraging a multi-level binary matching attention mechanism model.This algorithmic framework integrates subtitle text recognition and generation,subtitle segment extraction,a knowledge point extraction model,and a retrieval module.Experimental results show that,compared with the current knowledge point extraction model,the method of this system has achieved the optimal performance on some key indicators on multiple datasets such as Inspec,NUS,Krapivin,SemEval,KP20k,which fully proves the potential and value of this system in practical applications.

Large-scale Innovation Competition Evaluation Scheme Based on Multi-stage Evaluation

ZHANG Chang'en, CHENG Qing, SI Yuehang, HUANG Jincai

Computer Science. 2024, 51 (10): 86-93. doi:10.11896/jsjkx.240400063

Abstract

PDF(2624KB) ( 238 )

References | Related Articles | Metrics

Currently,large-scale innovation competitions are constantly emerging.The evaluation of such competitions has become an urgent problem to be solved due to subjective differences among experts and other reasons.This paper focuses on the research and design of evaluation schemes for large-scale innovation competitions.Through the analysis of the scoring results of the exis-ting competitions,the advantages and disadvantages of various evaluation schemes are comprehensively compared to find the best evaluation,so as to make the review process as programmed and efficient as possible,saving manpower and time resources.Firstly,the text constructs an expert allocation model to determine the “cross distribution” plan for reviewing experts,and uses an improved simulated annealing algorithm to solve the problem.Secondly,the text constructs a weighted model to compare four types of standard score calculation methods,and designs an improved standard score calculation method based on expert weights.Lastly,considering the correlation between large range and innovation,a range regression model is established to evaluate the model based on range.The proposed model and algorithm are widely applicable,and have important practical reference significance,and high application value.

Research and Implementation of Metaverse Educational Communication Technology Scheme Based on Edge Computing and WebRTC

NIU Guanchong, LIU Feixiang, YANG Wen, MIAO Qiguang, MAO Liang

Computer Science. 2024, 51 (10): 94-104. doi:10.11896/jsjkx.231200082

Abstract

PDF(4700KB) ( 264 )

References | Related Articles | Metrics

Metaverse education applies the meta-universe concept to the field of education,creating diverse learning environments and resources for students,and providing personalized learning recommendations and guidance.In the meta-universe scenario,performance can be enhanced and costs reduced through edge computing technologies,including the reduction of remote control latency and increased efficiency of terminal devices.To address challenges such as high device costs and significant communication latency in VR,AR education interaction scene,a communication technology solution based on edge computing and WebRTC is proposed.This solution aims to resolve congestion issues caused by a large number of users accessing educational scenarios,establishing the foundational communication infrastructure for meta-universe education.Leveraging a unity rendering platform deployed in the cloud and a WebRTC multi-user communication module deployed at the edge,the system achieves extremely low-latency audio-visual transmission,enhancing the transmission performance of mobile edge platforms.Ultimately,a cloud-edge collaboration audio-video transmission system is constructed.Through specific edge computing transmission strategies and end-to-end experimental validation,a meta-universe education system with low latency and high performance is realized.Experimental verification confirms the feasibility of metaverse education scenarios.

Study on Multi-task Student Emotion Recognition Methods Based on Facial Action Units

ZHANG Xiaoyun, ZHAO Hui

Computer Science. 2024, 51 (10): 105-111. doi:10.11896/jsjkx.240300059

Abstract

PDF(2146KB) ( 254 )

References | Related Articles | Metrics

With the rapid development of intelligent education,it has become a trend to use artificial intelligence to improve the quality and efficiency of education.The emotional state of students,who are at the center of education,has a crucial impact on educational effectiveness.In order to study students' emotions in depth,this paper collects students' learning videos in classroom scenarios,including two contexts of listening to lectures and group discussions,and builds a multi-task students' emotion database accordingly.The face serves as a direct outward manifestation of the internal emotional state,which shows a strong correlation between AU and emotions.Based on this,this paper proposes a multi-task learning-based student emotion recognition model Multi-SER.The model explores the association relationship between individual AUs and students' emotions by combining the two tasks of AU recognition and students' emotion recognition,thereby improving the performance of the model in students' emotion recognition.In the multi-task experiment,the Multi-SER model achieves an accuracy of 80.87% in emotion recognition,which improves the effect by 3.11% compared to the single emotion recognition task model SE-C3DNet+.The experimental results show that the performance of the model in categorizing various emotions is improved by mining the correlations between AUs and emotions through multi-task learning.

Eye Emotion Recognition and Visualization in Smart Classrooms Based on ConvNeXt

ZHANG Liguo, XU Xin, DONG Yuxin

Computer Science. 2024, 51 (10): 112-118. doi:10.11896/jsjkx.240400118

Abstract

PDF(3005KB) ( 311 )

References | Related Articles | Metrics

By leveraging facial expression recognition and emotion analysis,observers can understand learners' learning outcomes through the observed physical states.For instance,fluctuations in students' emotions displayed in the classroom can be used to discern their level of acceptance of new knowledge,facilitating a more convenient and intuitive understanding of students' confusion.However,in many cases,students' faces may be obstructed by learning materials,classmates in the front row,etc.,leading to inaccuracies in facial emotion recognition.Compared to the entire face,the eye region,as a core area of emotional expression,typically receives more attention from observers,and in the same classroom environment,the eyes are less likely to be obstructed.The eyes are one of the most important parts for displaying emotions,and changes in eye expressions during emotional fluctuations can provide more emotional information.Especially when a person is under external pressure and tend to suppress facial expressions,it is difficult to deceive with the gaze.Therefore,recognizing and analyzing complex eye expressions for emotions holds significant research value and challenges.To address this challenge,firstly,a dataset for classifying complex emotions in eye expressions is constructed,including five basic emotions,as well as defining five complex emotions.Secondly,a novel model is proposed to accurately classify emotions based on eye features extracted from input images in the dataset.Finally,a visualization method for emotion analysis based on eye recognition is introduced,which can analyze fluctuations in complex and basic emotions.This method provides a new solution for further eye-based emotion analysis.

Student Academic Performance Predictive Model Based on Dual-stream Deep Network

XIE Hui, ZHANG Pengyuan, DONG Zexiao, YANG Huiting, KANG Huan, HE Jiangshan, CHEN Xueli

Computer Science. 2024, 51 (10): 119-128. doi:10.11896/jsjkx.240300097

Abstract

PDF(1910KB) ( 212 )

References | Related Articles | Metrics

Blended teaching is one of the essential teaching methods with the development of information technology.Constructing a learning effect evaluation model is helpful to improve students' academic performance and helps teachers to better implement course teaching.However,a lack of evaluation models for the fusion of temporal and non-temporal behavioral data leads to an unsatisfactory evaluation effect.To meet the demand for predicting students' academic performance through learning behavior data,this study proposes a learning effect evaluation method that integrates expert perspective indicators to predict academic performance by constructing a dual-stream network that combines temporal behavior data and non-temporal behavior data in the learning process.In this paper,firstly,the Delphi method is used to analyze and process the course learning behavior data of students and establish an effective evaluation index system of learning behavior with universality;secondly,the Mann-Whitney U-test and the complex correlation analysis are used to analyze further and validate the evaluation indexes;and lastly,a dual-stream information fusion model,which combines temporal and non-temporal features,is established.The learning effect evaluation model is built,and the results of the mean absolute error(MAE)and root mean square error(RMSE) indexes are 4.16 and 5.29,respectively.This study indicates that combining expert perspectives for evaluation index selection and further fusing temporal and non-temporal behavioral features that for learning effect evaluation and prediction is rationality,accuracy,and effectiveness ,which provides a powerful help for the practical application of learning effect evaluation and prediction.

Research and Practice on “Five in One” Smart Teaching Model in the Course of CommunicationPrinciple

MA Yingjie, YANG Yatao, XIAO Song, LI Li

Computer Science. 2024, 51 (10): 129-134. doi:10.11896/jsjkx.240300007

Abstract

PDF(4119KB) ( 257 )

References | Related Articles | Metrics

In response to the strategic demand for digital transformation in education and teaching,and based on the characteristics of the course of communication principle,a new smart teaching model of the “Five in One” in the course of communication principle is proposed based on the OBE concept.The course of communication principle has outstanding theoretical,systematic,engineering,and practical characteristics.This paper starts from a global perspective and systematically designs the curriculum system,with the mission of cultivating morality,serving the Party,and cultivating talents in all aspects.Centered on student development,a “Five in One” smart teaching system has been constructed from five aspects:reconstructing teaching content,constructing a systematic system of ideological and political education in courses,implementing student project practices,creating smart classrooms,and implementing diverse evaluation mechanisms.Using digital and information-based teaching methods such as online and offline hybrid teaching in Rain Classroom,virtual simulation project teaching,and Bilibili micro course teaching,technology is used to empower teaching and create a smart classroom.Integrating ideological and political education into smart classroom teaching has a silent and subtle effect,triggering students to resonate with knowledge,emotions,and values.The new paradigm of “Five in One” smart education and teaching proposed in this paper can promote the deep integration of information technology,intelligent technology,and education and teaching,achieve intelligent teaching methods,ubiquitous teaching resources,real-time teaching comments,and diversified course assessments.

Recognition and Analysis of Teaching Behavior Based on Multi-scale GCN

LI Jia'nan, LI Ruiyi, ZHAO Zhifu, SONG Juan, HAN Jialong, ZHU Tong

Computer Science. 2024, 51 (10): 135-143. doi:10.11896/jsjkx.240400089

Abstract

PDF(2224KB) ( 274 )

References | Related Articles | Metrics

In the field of education,classroom teaching evaluation stands as a pivotal element in enhancing teaching quality.With the widespread adoption of digital education,the quest for an intelligent evaluation method becomes increasingly crucial.Therefore,this paper proposes a novel method based on skeleton action recognition and lagged sequence analysis,aiming to more accurately capture and analyze teachers' teaching behaviors while reducing manpower consumption and diminishing the subjectivity of teaching evaluations.Firstly,a multi-scale feature graph convolutional network is proposed and applied to analyze teacher classroom behaviors.This network utilizes a multi-scale semantic feature fusion module to capture features at two scales,skeleton points,and body parts,in the spatial dimension.In the temporal dimension,a multi-scale temporal feature extraction module is employed to extract temporal features of skeleton data from both global and local perspectives.Subsequently,a dataset for analyzing teachers' classroom behaviors is constructed,and the effectiveness of the proposed method is validated on this dataset.Finally,leveraging the proposed skeleton action recognition model and lagged sequence analysis,a system for recognizing and analyzing teaching behaviors is developed.The proposed method demonstrates significant advantages in classroom behavior recognition and analysis when applied to various classroom teaching scenarios.

Virtual Interactive Teaching Skills Training Mode for Normal Students in Educational MetaversePerspective

YANG Kaifang, LI Junchi, XU Yan, GONG Yanchao

Computer Science. 2024, 51 (10): 144-152. doi:10.11896/jsjkx.240400120

Abstract

PDF(4914KB) ( 293 )

References | Related Articles | Metrics

With the rapid development of information technology,emphasizing digital education reform in the process of cultivating outstanding teachers has become an inevitable requirement for the high-quality development of education in the future.In recent years,the concept of metaverse and educational metaverse has become a research hotspot,which brings new reform ideas to the training methods of teaching skills for normal students.The traditional training of teaching skills for normal students mainly focuses on classroom practices,and is limited by practicing time and spaces.This inevitably resulting in lack of a personalized,a specialized,and a targeted training for normal students.With the development of information technology,“Internet＋” training mobile training,and other methods are proposed,however,the training process is not highly immersive and interactive.Therefore,this paper integrates the concept of the educational metaverse,and proposes a novel virtual interactive teaching skill training mode for normal students.Taking the training of body language skills and introduction skills as examples,it introduces the construction and practical process of the proposed virtual interactive teaching skill training mode for normal students from the perspective of the educational metaverse,analyzes the practical effect,and verifies the effectiveness of the virtual interactive teaching skill trai-ning mode for normal students.

Learning Path Recommendation Method Based on Feature Similarity and Jaccard Median

YANG Pengfei, WANG Shuqi, HUANG Jiayang, ZHANG Wenjuan, WANG Quan, ZHONG Haodi

Computer Science. 2024, 51 (10): 153-161. doi:10.11896/jsjkx.240400111

Abstract

PDF(3466KB) ( 259 )

References | Related Articles | Metrics

The advancement of the new college entrance examination has prompted more and more colleges to convert their enrollment mode from professional enrollment to enrollment in general categories.However,relevant studies indicate that there is a lack of rationality in students' choices when it comes to major shunts.How to break the situation of “cold majors and hot majors” caused by the imbalance of major selection has become the core problem faced by large types of training models.A learning path recommendation method based on feature similarity and Jaccard median(CFSJM) is proposed in this paper,aiming to provide direction navigation and learning path recommendations for students when choosing their majors.The method utilizes Node2vec to learn the interactions between students and knowledge points to obtain a feature representation of student nodes.A linear regression model is trained to predict the students' major direction,and a learning path candidate set is generated based on feature similarity,which in turn introduces the Jaccard median theory to generate learning paths.Experimental results show that the accuracy of CFSJM in the offline teaching data is better than that of the existing methods,which provides a new idea to give full play to the advantages of enrollment in general categories in cultivating innovative talents and improving the quality of university education.

Prerequisite Relation Information Enhanced Relation Prediction Method for Course KnowledgeGraph

YANG Jiaqi, HE Chaobo, GUAN Quanlong, LIN Xiaofan, LIANG Zhuoming, LUO Huiqiong

Computer Science. 2024, 51 (10): 162-169. doi:10.11896/jsjkx.240400090

Abstract

PDF(3046KB) ( 256 )

References | Related Articles | Metrics

A large amount of course knowledge graphs have played a crucial role in intelligent teaching applications such as automatic Q&A,learning path planning,and learning resource recommendation.However,the incompleteness issue caused by missing entity relations significantly reduces their application value.Relation prediction is the primary means of automatically completing the missing relations in course knowledge graphs,but existing methods only directly use sparse topology information and fail to exploit and enhance the prediction performance by further using its unique prerequisite relation information.To address this pro-blem,a course knowledge graph relation prediction method,prerequisite relation information enhanced relation prediction(PRIERP),is proposed.This method first designs a prerequisite relation information extraction mechanism based on semantic path computation.Then,it constructs dual views based on topology information and prerequisite relation information,and designs a directed graph Transformer to learn the low-dimentional representation of the course knowledge graph from the dual views.Finally,an end-to-end relation prediction is achieved based on a multi-layer perceptron classification model.Experiments are conducted on two typical course knowledge graphs HhsMath and ML.The results demonstrate that PRIERP outperforms other representative methods.In HhsMath,PRIERP achieves at least 2.43%,5.93%,4.73% and 1.72% improvements in terms of MRR,Hits@1,Hits@3,and Hits@10 metrics,respectively.Furthermore,the analysis of typical cases in relation prediction also confirms its effectiveness.

Speed-Accuracy Tradeoff-based Deep Cognitive Diagnostic Model

CHENG Yan, ZHOU Ziwei, MA Mingyu, LIN Qinglong, ZHAN Yongxin, WAN Lingfeng

Computer Science. 2024, 51 (10): 170-177. doi:10.11896/jsjkx.240300121

Abstract

PDF(2294KB) ( 242 )

References | Related Articles | Metrics

In intelligent education,cognitive diagnosis analyzes students' learning behavior data to understand their cognitive state.Existing cognitive diagnostic models based on deep learning methods assume by default that students have enough reaction time to fully exert the level of knowledge mastery during the response process,and do not consider the impact of the trade-off strategy between the speed and accuracy of student's response during the response process on the exertion of the level of know-ledge mastery.Aiming at the above problem,a deep cognitive diagnostic model based on speed-accuracy trade-off is proposed,which firstly constructs a cognitive style fuzzy set to explain the students' trade-off strategy,and then simulates the speed-accuracy trade-off relationship in the process of the learners' response through the dynamic logistic regression function,so as to rea-lize the differentiated diagnosis of the students' theoretically highest level of knowledge mastery from the level of knowledge mastery they have played out in the actual response.In addition,the reaction time attribute and exercise type attribute are introduced to more accurately characterize the topic parameters in the cognitive diagnostic interaction function.Numerous experiments show that the model not only improves the accuracy by 2.58%,2.86%,and 5.18% compared to similar optimal models on the three publicly available datasets,but also provides a superior explanation of the prediction results at the level of response time.

Study on Building Business-oriented Resource On-demand Resolution Model

LIU Yao, QIN Xun, LIU Tianji

Computer Science. 2024, 51 (10): 178-186. doi:10.11896/jsjkx.230800191

Abstract

PDF(2409KB) ( 226 )

References | Related Articles | Metrics

To address the issue of re-analyzing and repeating development of natural language processing tools and resource ana-lysis plugins when new requirements arise during project development,this paper proposes a business-oriented on-demand resource analysis solution.Firstly,a demand-driven resource analysis method from requirement to code is proposed,focusing on the construction of a demand concept indexing model for the requirement text itself.The constructed demand concept indexing model outperforms other classification models in terms of accuracy,recall,and F1 score.Secondly,this paper establishes a mapping mechanism from requirement text to code library categories based on the correlation between requirement text and code.For the mapping results,the precison@K is used as an evaluation metric,with an ultimate accuracy rate of 60%,demonstrating a certain practical value.In summary,this paper explores a set of key technologies for on-demand resource analysis with demand parsing capabilities and implements the correlation between requirements and code,covering the entire process from requirement text classification,code library classification,code library retrieval to plugin generation.The proposed method forms a complete business loop of “requirement-code-plugin-analysis” and experimentally verifies to be effective for on-demand resource analysis.Compared to existing large language models for business requirement analysis and code generation,this method focuses on the implementation of the full process of plugin code reuse within specific business domains,containing business characteristics.

Data Mining and Information Service for Open Collaboration Digital Ecosystem

XIA Xiaoya, ZHAO Shengyu, HAN Fanyu, BI Fenglin, WANG Wei, ZHOU Xuan, ZHOU Aoying

Computer Science. 2024, 51 (10): 187-195. doi:10.11896/jsjkx.230900071

Abstract

PDF(1984KB) ( 218 )

References | Related Articles | Metrics

Large-scale development and proliferation of open source software has constructed an ecosystem for open source deve-lopment and collaboration.Within this system,individuals and organizations collaboratively develop high-quality software that is accessible to all.Social collaboration platforms,represented by GitHub,have further facilitated large-scale,distributed,and fine-grained code collaboration and technical socialization.Countless developers submit code,review code,report bugs,or propose new feature requests on these platforms every day.This results in a vast amount of behavioral data from the fully open collaborative development process,which holds immense value.This paper designs and implements a one-stop data mining system for the open source collaboration digital ecosystem,named OpenDigger.Its goal is to build data infrastructure in the open source field and promote the continuous development of the open source ecosystem.OpenDigger system consists primarily of data collection module,storage module,tag data module,and information service module.It is built upon an OLAP columnar database and a graph database.The system continuously collects data from multiple sources within the open-source ecosystem and provides various types of open-source information services to different user groups through a unified interface.Additionally,OpenDigger mines key information from the open-source digital ecosystem through the perspective of collaborative relationship networks.Compared to traditional statistical indicators,the collaborative network perspective better illustrates the association characteristics between open-source projects and developers.

Keyword Sensitive Fuzzing Method for Embedded Device Firmware

SI Jianpeng, HONG Zheng, ZHOU Zhenji, CHEN Qian, LI Tao

Computer Science. 2024, 51 (10): 196-207. doi:10.11896/jsjkx.230700068

Abstract

PDF(3213KB) ( 226 )

References | Related Articles | Metrics

The firmware of most embedded devices provides a Web interface,which is convenient for the users to configure and manage the devices.However,the security problems of these Web interfaces usually bring challenges to the security of embedded devices.However,the existing vulnerability detection methods for Web interfaces in embedded device firmware have high false positive rates.This paper proposes a keyword-sensitive embedded device fuzzing method KS-Fuzz(keyword sensitive fuzzing),which efficiently performs fuzzing in the processing logic of the Web interface in the embedded device firmware.The proposed method generates high-quality test cases through the association analysis of front-end and back-end files,and records the refe-rences of keywords in the target device's back-end files to front-end files during the fuzzing process,to guide the direction of test case mutation,and improve the fuzzing coverage.In this paper,we use KS-Fuzz to test embedded devices of major brands to eva-luate the fuzzing ability of KS-Fuzz,and compare KS-Fuzz with existing vulnerability mining methods,such as SaTC,IOTScope,and FirmFuzz.The results show that by analyzing the correlation of front-end and back-end files,KS-Fuzz can quickly traverse the functional interfaces of the target devices and discover vulnerabilities effectively.

Robust Binary Program Debloating

DING Duo, SUN Cong, ZHENG Tao

Computer Science. 2024, 51 (10): 208-217. doi:10.11896/jsjkx.230700008

Abstract

PDF(2776KB) ( 244 )

References | Related Articles | Metrics

The frequently used functionalities usually constitute a small portion of applications' functionalities.The redundant code for rarely used functionalities raises the attack surface of the applications,thus causing the potential risk of code reuse attacks.Binary program debloating can identify and remove the redundant code based on the binary analysis of the application,so as to reduce the attack surface.The state-of-the-art binary program debloating approach relies on artificially crafted inputs to derive the initial control flows.It uses heuristics to extend the binary control-flow graph for debloating.Such an approach has limited robustness and scalability.This paper proposes and implements a robust binary program debloating approach(RBdeb).It uses black-box fuzzing to derive highly-robust valid execution traces of the binary,and categorizes similar library functions automatically based on the graph isomorphism algorithm.The proposed path discovery algorithm extends the binary control flows with the classified library function calls from the control-flow sub-graph of the initial execution traces and generates the robust binary file as the debloating result.Experimental results demonstrate that RBdeb has higher path coverage and debloated binary robustness than the state-of-the-art approaches.The path discovery algorithm and library function categorization are more scalable.RBdeb can effectively debloat large real-world applications.

Risk Assessment Model for Industrial Chain Based on Neighbor Sampling and GraphAttention Mechanism

SUN Pengzhao, BI Kejun, TANG Chao, LI Dongfen, YING Shi, WANG Ruijin

Computer Science. 2024, 51 (10): 218-226. doi:10.11896/jsjkx.230900145

Abstract

PDF(2861KB) ( 212 )

References | Related Articles | Metrics

Risk assessment is an important way to improve the resilience of the industrial chain and also an effective method to reduce the instability of the industrial chain.However,existing research on risk assessment is based on supply chain structure and neglects other factors,which can not accurately depict the correlation between upstream and downstream nodes in the industrial chain,resulting in biased evaluation results.In response to the above issues,considering the interconnected nature of various nodes within the industry chain,diverse risk situations,and the existence of risk transmission,this paper proposes an industry chain risk assessment model based on graph attention mechanism and neighbor sampling(GANS).Firstly,a heterogeneous graph of the industrial chain is constructed,using “product-company” and “product-product” to depict the correlation between nodes in the industrial chain,and financial data and other data features are extracted from the industrial chain as nodes' data features.Se-condly,a company relationship graph generation module based on meta paths and company investment and financing association rules is proposed to achieve efficient transformation of company node relationships and efficient learning of structural features in the industrial chain.Next,an industry chain risk assessment module based on neighbor sampling and graph attention mechanism is designed for various generated company graphs.The features of node neighbors are randomly sampled and aggregated,and attention mechanism is used to adaptively aggregate node features based on multiple company graphs.Through the classifier,node-level risk assessment is realized.Finally,risk assessment of the industrial chain is conducted based on the risk level and structural features of nodes.Experiments show that GANS outperforms existing models in terms of accuracy and F1 scores on real indu-strial chain datasets.These results demonstrate the accuracy and effectiveness of GANS in implementing industrial chain risk assessment.

Structural Influence and Label Conflict Aware Based Graph Curriculum Learning Approach

LIU Zulong, CHEN Kejia

Computer Science. 2024, 51 (10): 227-233. doi:10.11896/jsjkx.230800167

Abstract

PDF(2810KB) ( 224 )

References | Related Articles | Metrics

In recent years,graph neural networks(GNNs) have emerged as a prominent research area in the field of graph lear-ning.Leveraging the message passing mechanism,GNNs have showcased remarkable performance across diverse graph-based tasks.However,most existing GNNs methods assume uniform training difficulty across all nodes,disregarding the significant va-riability in the importance and contributions of different nodes.To address this problem,this paper proposes a structural influence and label conflict aware graph curriculum learning method(SILC-GCL),which takes the training difficulty of nodes into conside-ration.To begin with,a difficulty measure is designed through both the graph structure and node label semantics,calculating the PageRank value of nodes and the label conflict degree between nodes and their neighbors.Subsequently,a training scheduler is employed to select nodes with appropriate training difficulty at each training stage and then generate a sequence of training nodes from easy to difficult.Finally,SILC-GCL is trained based on backbone GNNs models.Experimental results of node classification on six benchmark datasets verify the effectiveness of SILC-GCL.

NLGAE:A Graph Autoencoder Model Based on Improved Network Structure and Loss Functionfor Node Classification Task

LIAO Bin, ZHANG Tao, YU Jiong, LI Min

Computer Science. 2024, 51 (10): 234-246. doi:10.11896/jsjkx.230700122

Abstract

PDF(3898KB) ( 258 )

References | Related Articles | Metrics

The universally accepted technique to address the issues of computational complexity and high spatial complexity of adjacency matrix due to non-Euclidean spatiality of graph data is to use graph embedding methods to map high-dimensional heterogeneous information,such as graph topology and node attributes,to dense vector space.In this paper,based on the analysis of the problems of the classical graph auto-encoder model GAE(graph auto-encoder) and VGAE(variational graph auto-encoder),we try to improve the graph embedding method based on graph auto-encoder from three aspects:encoder,decoder and loss function,and propose a graph auto-encoder model NLGAE based on the improved network structure and loss function.First,in the model structure design,on the one hand,the stacked graph convolutional layers in the encoder are inverted to solve the problem of lack of flexibility and insufficient expressiveness of the non-reference decoder in GAE and VGAE,on the other hand,the graph convolutional network GAT is introduced to solve the problem of solidifying the weight coefficients between nodes by introducing the attention mechanism.Second,both the graph structure and the node feature information could be taken into account by the redesigned loss function.The comparative experimental results show that,as an unsupervised model,the proposed NLGAE can learn high-quality node embedding features and outperform not only traditional unsupervised models DeepWalk,GAE,GrpahMAE,GATE,etc.in node classification tasks,but also supervised graph neural network models such as GAT and GCN in the case of selecting an appropriate classification model.

All-chain Sets Mining Algorithm for Multi-scale Nearest Time Series

WANG Shaopeng, FENG Chunkai

Computer Science. 2024, 51 (10): 247-260. doi:10.11896/jsjkx.230800146

Abstract

PDF(4824KB) ( 250 )

References | Related Articles | Metrics

Mining all-chain set in the time series is an emerging area.To the best of our knowledge,no method has been proposed to mining all-chain sets over multi-scale nearest time series.In this paper,the problem of mining all-chain sets over multi-scale nearest time series is focused.The mining problem of all-chain sets over multi-scale nearest time series is studied,and a mining algorithm with incremental computation characteristics is proposed on the basis of the existing LRSTOMP and ALLC algorithms,MTSC(mining time series all-chain sets over multi-scale nearest time series).The MTSC algorithm uses the LRSTOMP and ALLC algorithms sequentially to process the content of the 1st nearest time series member to obtain the mining results of all-chain sets over this member,while keeping the PL and PR structures associated with this member.Starting from the 2nd nearest time series member,the LRSTOMP process in the MTSC algorithm only needs to deal with the additions of the current nearest time series member with respect to the previous nearest time series member,and further combining the PL and PR on the pre-vious nearest time series member can incrementally obtain the structure of the PL and PR on the current nearest time series member,and based on which the ALLC algorithm is used to get the all-chain set mining result on that member.Compared to the Naive way using LRSTOMP and ALLC algorithms to process the content of each recent time series member,the MTSC algorithm avoids repetitive computation on all data by utilizing the idea of incremental computation,which improves the execution speed of the algorithm and has better time efficiency.Simulation experiments based on the common data samples Penguin and TiltABP verify the effectiveness of the proposed algorithm,and the experiment results show that the results of the MTSC algorithm are completely consistent with that of the Naive algorithm,and the MTSC algorithm can achieve 80%~ 88.3% improvement in time efficiency for the above data samples with an increase in space overhead of 1.1% ~ 9.7%.

Research Progress in Industrial Defect Detection Based on Deep Learning

HONG Jingshan, ZHU Yingdan, SONG Kangkang, LYU Dongxi, CHEN Mingda, HU Haigen

Computer Science. 2024, 51 (10): 261-275. doi:10.11896/jsjkx.230800158

Abstract

PDF(3518KB) ( 322 )

References | Related Articles | Metrics

Machine vision technology based on deep learning has important application value in industrial defect detection,which can significantly improve detection quality,efficiency and reduce labor costs compared to traditional methods.By collecting the research and application information of deep learning in defect detection in recent years,the difficulties and related solutions are summarized,and the problems are divided into two aspects:the problem of establishing defect datasets and the selection of detection models.First,at the data aspect,aiming at the problems of few samples of defects,data labeling,and low quality of data imaging,this paper correspondingly analyzes the applications of small sample learning,unsupervised,semi-supervised,self-supervised and weak-supervised learning,data augmentation,image enhancement and image translation.Then,in the selection of neural network models,according to the different types of models,they are divided into three categories:CNN based,Transformer based,and mixture model for discussion.According to different detection requirements,they are divided into three types of models:classification,detection,and segmentation.In addition,the design methods of lightweight models are summarized.Finally,the future development direction is discussed and prospected.

Review of Quality Control Algorithms for Pathological Slides Based on Deep Learning

DING Weilong, LIU Jinlong, ZHU Wei, LIAO Wanyin

Computer Science. 2024, 51 (10): 276-286. doi:10.11896/jsjkx.231000167

Abstract

PDF(2015KB) ( 255 )

References | Related Articles | Metrics

Pathological sections are an important basis for pathologists to diagnose and analyze the course and prognosis of tumor patients.However,due to the low degree of automation in the preparation process of slides,both human operation and equipment noise will reduce the quality of the slides,thereby affecting diagnosis.Currently,the quality control of pathological slides mainly uses manual sampling inspection,which has the characteristics of high work intensity and long working hours,and can easily lead to evaluation bias due to visual fatigue.Quality control of pathological slides using deep learning technology attracts attention from the medical and engineering communities and makes certain progress.This paper reviews the research status in this field.First,the production and digitization process of pathological slides is briefly introduced,and the difficulties and challenges in qua-lity control work are analyzed.Then,the existing work related to the quality control of pathological slides is analyzed and summarized,and the method theory and application status of the existing work are reviewed from aspects such as staining standardization,focus quality assessment,artifact detection,image repair and reconstruction.Finally,the possible future research hotspots in this field are prospected.

Infrared Dim and Small Target Detection Based on Cross-domain Migration of Visible Light andInfrared

XUE Ruxiang, WEI Junjie, ZHOU Huawei, YANG Hai, WANG Zhe

Computer Science. 2024, 51 (10): 287-294. doi:10.11896/jsjkx.230800013

Abstract

PDF(2986KB) ( 288 )

References | Related Articles | Metrics

The task of infrared dim and small target detection is one of the key research contents in the field of infrared detection.However,due to the particularity of its application scenarios,the data containing infrared dim and small targets is rare,and often not fully labeled,which poses challenges and difficulties for data-driven deep learning object detection models.In order to solve the problems of limited datasets and lack of label information,an infrared dim and small target detection model based on cross-domain migration of visible light and infrared is proposed to migrate the more abundant visible light domain supervision information to the infrared domain,so as to achieve unsupervised training in the infrared domain.First,a channel augmentation data proces-sing method is designed on the basis of YOLOv5,utilizing low-cost channel separation techniques to convert visible light images into infrared like images,reducing the modal differences between the visible and infrared domains.Then,a multi-scale domain adaptive module is constructed,and the features of different scales extracted by the backbone network are used in the way of adversarial training.Domain confusion is performed in the feature space to reduce the impact of domain shift and improve the detection performance of dim and small target detection.Experimental results show that the improved model by the proposed method can improve the average detection precision compared to various versions of the YOLOv5 original model.Compared with other existing unsupervised domain adaptive target detection algorithms,the proposed method is obviously superior in the detection accuracy of small infrared targets.

Eye Gaze Estimation Network Based on Class Attention

XU Jinlong, DONG Mingrui, LI Yingying, LIU Yanqing, HAN Lin

Computer Science. 2024, 51 (10): 295-301. doi:10.11896/jsjkx.230900094

Abstract

PDF(2317KB) ( 224 )

References | Related Articles | Metrics

In recent years,eye gaze estimation has attracted widespread attention.The gaze estimation method based on RGB appearance uses ordinary cameras and deep learning for gaze estimation,avoiding the use of expensive infrared devices like commercial eye trackers,providing the possibility for more accurate and cost-effective eye gaze estimation.However,due to the presence of various features unrelated to gaze,such as lighting intensity and skin color,in RGB appearance images,these irrelevant features can cause interference in the deep learning regression process,thereby affecting the accuracy of gaze estimation.In response to the above issues,this paper proposes a new architecture called class attention network(CA-Net),which includes three different class attention modules:channel,scale,and eye.Through these class attention modules,different types of attention encoding can be extracted and fused,thereby reducing the weight of gaze independent features.Extensive experiments on the GazeCapture dataset show that,compared to the state-of-the-art method,CA-Net can improve gaze estimation accuracy by approximately 0.6% and 7.4% on mobile phones and tablets,respectively,in RGB based gaze estimation methods.

Order-adaptive Multi-hypothesis Reconstruction for Heterogeneous Image Compressive Sensing

ZHENG Yongxian, LIU Hao, YAN Shuai, CHEN Genlong

Computer Science. 2024, 51 (10): 302-310. doi:10.11896/jsjkx.230800156

Abstract

PDF(2554KB) ( 214 )

References | Related Articles | Metrics

The arrival of the big data era poses challenges for processing and transmitting large amounts of image data.The compressive sensing technology and related algorithms have solved some of these problems to a certain extent.However,existing compressive sensing algorithms still have problems when adapting to heterogeneous image sets.Therefore,it is necessary to design a highly generalized compressive sensing reconstruction algorithm for such image sets.In this paper,an order-adaptive multi-hypothesis reconstruction algorithm is proposed according to a multih-ypothesis prediction mechanism with high generalization.The proposed algorithm preprocesses each block using a window-adaptive linear predictor and changes the size of the multi-hypothesis searching window according to the correlation index obtained from preprocessing.The prediction blocks within the searching window are sorted according to block-wise similarity and different numbers of highly similar prediction blocks are selected from the adaptive searching window for the reconstructed image of multi-hypothesis prediction.Experiments are conducted on a natural image set and two heterogeneous image sets of X-ray chest and brain MRI.At different sampling rates,many experiments and analyses are carried out by comparing the traditional multi-hypothesis compressive sensing reconstruction algorithm and two recent algorithms of multi-hypothesis prediction.The experimental results show a good performance improvement of the proposed algorithm compared to the traditional multihypothesis compressive sensing reconstruction algorithm.On the natural image set,the proposed algorithm maintains a certain recovery quality and achieves an average runtime decrease of 17.5% and 28.7% respectively,compared to two recently proposed algorithms.As compared to two recent proposed algorithms:on the X-ray chest image set,the average PSNR value of proposed algorithm increases by 1.16dB and 1.43dB,and the average runtime decreases by 36.1% and 21.5%,respectively.On the brain MRI image set,the average PSNR value increases by 1.64dB and 1.97dB,and the average runtime decreases by 28.6% and 26.1%,respectively.Overall,the proposed algorithm has low computational complexity and high recovery quality with better tradeoff performance.

Infrared and Visible Deep Unfolding Image Fusion Network Based on Joint Enhancement ImagePairs

YUAN Tianhui, GAN Zongliang

Computer Science. 2024, 51 (10): 311-319. doi:10.11896/jsjkx.230800069

Abstract

PDF(5607KB) ( 243 )

References | Related Articles | Metrics

Under unfavorable circumstances,the fused image of the infrared and visible images sometimes suffers from low brightness and insufficient details.Therefore,a novel infrared and visible deep unfolding image fusion network based on joint enhancement image pairs is proposed.To increase input information,both the original infrared/visible image pair and their enhancement pair are used as deep network's input.Firstly,an iterative residual unfolding convolutional network based on deep residual unfolding module is developed to obtain the background features or detail features according to different initialization network parameters.Then,concatenate operation and up-down sampling pair are introduced to the convolutional feature fusion network,where features of the corresponding enhancement image pairs can be added to fusion task and the discrepant features of raw images are maximumly retained.Meanwhile,the loss function is optimized to obtain better results.Numerous experiments on multiple datasets demonstrate that the proposed method can get competitive fusion images both in terms of subjective evaluation and objective metrics,and have better performance under low light environments.

Mechanical Fault Diagnosis Under Variable Working Conditions Based on Sharpness AwarenessReinforced Convolutional Neural Network

FAN Jiayuan, XU Desheng, LUO Lingkun, HU Shiqiang

Computer Science. 2024, 51 (10): 320-329. doi:10.11896/jsjkx.230900139

Abstract

PDF(2909KB) ( 250 )

References | Related Articles | Metrics

Traditional deep transfer learning networks have effectively addressed the challenges arising from the asymmetry introduced by cross-domain data distributions in variable operational scenarios.It is achieved by leveraging knowledge learned from labeled fault data and applying it to the task of diagnosing unlabeled fault data collected under varying conditions.However,the inclusion of knowledge transfer modules has added complexity to the deep network's structure,resulting in a more intricate loss landscape.This,in turn,presents challenges for optimization.Traditional methods often struggle to navigate the sharpness of this loss landscape,leading to the model's parameters getting stuck in local minima characterized by high sharpness.This hinders model generalization and reduces accuracy.To tackle this challenge,this paper proposes the sharpness awareness reinforced con-volutional neural network(SA-CNN).This approach involves a joint optimization of the loss function and its flatness by assessing sharpness within a specified range.This process steers the fault diagnosis model parameters away from regions of high sharpness,ultimately improving model generalization.Extensive experiments on established mechanical fault diagnosis datasets demonstrate that,compared to traditional deep transfer learning-based fault diagnosis models,the proposed SA-CNN significantly enhances the performance of bearing fault diagnosis under varying working conditions.

Decision Algorithms for Reversibility of One-dimensional Non-linear Cellular Automata Under Null Boundary Conditions

MA Junchi, CHEN Weilin, WANG Chen, LIN Defu, WANG Chao

Computer Science. 2024, 51 (10): 330-336. doi:10.11896/jsjkx.240100207

Abstract

PDF(1884KB) ( 198 )

References | Related Articles | Metrics

The property of reversibility is quite meaningful for the classic theoretical computer science model,cellular automata.For the reversibility problem for a CA under null boundary conditions,while linear rules have been extensively studied,the non-linear rules have so far been rarely explored.The paper investigates the reversibility problem of general one-dimensional CA on a finite field $\mathbb{Z}_{p}$,and proposes an approach to optimize the Amoroso's infinite CA surjectivity detection algorithm.Based on this,this paper proposes an algorithm for determining the invertibility of one-dimensional CA under zero boundary conditions,including an algorithm for determining the strict invertibility of one-dimensional CA under zero boundary conditions,and an algorithm for calculating the invertibility function of one-dimensional CA under zero boundary conditions based on barrel chain.These decision algorithms work for not only linear rules but also non-linear rules.In addition,it has been confirmed that the reversibility function always has a period,and its periodicity is related to the periodicity of the corresponding bucket chain.Some of the experiment results of reversible CA are presented in this paper,complementing and validating the theoretical aspects,and thereby further supporting the research conclusions of this paper.

Document-level Relation Extraction Based on Multi-relation View Axial Attention

WU Hao, ZHOU Gang, LU Jicang, LIU Hongbo, CHEN Jing

Computer Science. 2024, 51 (10): 337-343. doi:10.11896/jsjkx.230800033

Abstract

PDF(2715KB) ( 245 )

References | Related Articles | Metrics

Document-level relationship extraction aims to extract relationships between multiple entities from documents.To address the limited multi-hop reasoning capacity of existing methods for establishing connections between entities with different relationship types,this paper propose a document-level relationship extraction model based on multi-relation view axial attention.The model will construct a multi-view adjacency matrix based on the relationship types between entities,and use it to perform multi-hop reasoning.In order to evaluate the proposed model's performance,two benchmark datasets for document-level relationship extraction,namely GDA and DocRED are used in this study.The experimental results demonstrate that the F1 metric achieves 85.7% on the biological dataset GDA,significantly surpassing the baseline model's performance.Moreover,the proposed model proves effective in capturing the multi-hop relationships among entities in the DocRED dataset.

Generation of Contributions of Scientific Paper Based on Multi-step Sentence Selecting-and-Rewriting Model

XU Xianzhe, CHEN Jingqiang

Computer Science. 2024, 51 (10): 344-350. doi:10.11896/jsjkx.230800080

Abstract

PDF(2782KB) ( 216 )

References | Related Articles | Metrics

There has been a significant surge in the number of scientific papers published in recent years,which makes it challen-ging for researchers to keep up with the latest advancements in their fields.To stay updated,researchers often rely on reading the contributions section of papers,which serves as a concise summary of the key research findings.However,it is not uncommon for authors to inadequately present the innovative content of their articles,making it difficult for readers to quickly grasp the essence of the research.To address this issue,we propose a novel task of contribution summarization to automatically generate contribution summaries of scientific papers.One of the challenges of this task is the lack of relevant datasets.Therefore,we construct a scientific contribution summarization corpus(SCSC).Another issue lies in the fact that currently available abstractive or extractive models tend to suffer from either excessive redundancy or a lack of coherence between sentences.To meet the demand of ge-nerating concise and high-quality contribution sentences,we present MSSRsum,a multi-step sentence selecting-and-rewriting model.Experiments show that the proposed model outperforms baselines on SCSC and arXiv datasets.

Chemical-induced Disease Relation Extraction:Graph Reasoning Method Based on Evidence Focusing

ZHOU Xueyang, FU Qiming, CHEN Jianping, LU You, WANG Yunzhe

Computer Science. 2024, 51 (10): 351-361. doi:10.11896/jsjkx.230800111

Abstract

PDF(3438KB) ( 236 )

References | Related Articles | Metrics

To address the problem of existing methods focusing too much on global information while neglecting a small amount of evidence clues and local mention interactions when mining the interaction between chemicals and diseases,a mention level document-level relation extraction method based on evidence focusing(EF-MUnet) is proposed.This method first models mention features based on context aware strategies and captures local interactions between adjacent mentions using two-dimensional convolution network.Secondly,to avoid irrelevant context interference,two evidence focusing strategies ATT-EF and RL-EF are proposed.The former uses similarity as a measure of evidence clues,while the latter uses reinforcement learning to unsupervised learn the optimal evidence extraction policy with the help of delayed feedback.Finally,U-net networks are used to capture global features at the entity level and fully explore semantic relationships.Experimental results show that compared with existing me-thods,EF-MUnet's F1 score improves by 9.7% on the biomedical dataset CDR,and it has more advantages in extracting inter-sentence relations.In addition,EF-MUnet achieves the highest accuracy of 98.6% on the dataset DMI for extracting interactions between drug and mutation,proving that it is an effective biomedical relation extraction method with good generalization ability.

Open FaaS-based Multi-edge Management Framework

LIN Jingfeng, LI Ming, CHEN Xing, MO Yuchang

Computer Science. 2024, 51 (10): 362-371. doi:10.11896/jsjkx.230800203

Abstract

PDF(4134KB) ( 263 )

References | Related Articles | Metrics

Mobile edge computing(MEC) is a cutting-edge technology that utilizes the computing power provided by edge nodes close to mobile devices to improve performance.Some existing state-of-the-art computation offloading mechanisms support dynamic offloading of applications at the function granularity.Function as a service(FaaS),as a typical paradigm of the serverless architecture,enables a new way of building and scaling applications at function granularity.FaaS provides ideal resource elasticity compared to traditional approaches.OpenFaaS,as a popular open-source FaaS project,enables a good foundation for building FaaS platforms.Integrating advanced computation offload mechanism with FaaS solution(OpenFaaS) is meaningful and challenging.To this end,a multi-edge management framework based on OpenFaaS is designed and implemented in the paper.The framework rea-lizes the construction and status management of OpenFaaS on multiple edges.Also,for the functions that need to be deployed,they are reconstructed and then are deployed to OpenFaaS.Furthermore,at runtime,the framework is capable of flexibly scheduling function execution among multiple OpenFaaS instances.The framework is evaluated for 5 real-world Java intelligent applications.Results show that,compared to local invocation,the proposed framework saves response time by 10.49%～49.36% on average and the framework can effectively manage multiple edges.

Study on Stream Data Authorization Revocation Scheme Based on Smart Contracts

MEN Ruirui, JIA Hongyong, DU Jinru

Computer Science. 2024, 51 (10): 372-379. doi:10.11896/jsjkx.230700094

Abstract

PDF(2577KB) ( 231 )

References | Related Articles | Metrics

IoT devices and services encrypt real-time generated stream data for outsourced storage,and authorize users through access control.When the user's identity or permissions change,authorization to the user needs to be revoked.Existing revocation schemes have problems of frequent key updates and re-encrypted ciphertext,resulting in low revocation efficiency,insufficient flexibility,difficulty in achieving real-time revocation,and the risk of data leakage.In order to solve the real-time authorization revocation in the outsourcing storage scenario of streaming data,a decentralized authorization revocation scheme based on smart contracts is proposed.Under the IoT architecture combined with edge computing and blockchain,the streaming data is divided into blocks according to time intervals,and a large number of unique keys corresponding to the blocks are generated using the HASH tree,and the partitioned data is symmetrically encrypted.The tree nodes create access tokens and share them through proxy re-encryption technology,to implement modifiable access policies and efficient dynamic data sharing.By utilizing smart contract technology to create access control lists and misconduct lists,users are subjected to scheduled and immediate revocation operations,achieving decentralized real-time authorization revocation.Through security analysis and simulation experiments,it has been proven that this scheme provides better security,functionality,communication,and computing costs compared to other rela-ted research schemes,and is more effective.

SSPN-RA:Security Integration Risk Assessment Method for ICS Based on SS-petri Net

MA Zigang, MA Rongkuan, LI Beibei, XIE Yaobin, WEI Qiang, PENG Minwei

Computer Science. 2024, 51 (10): 380-390. doi:10.11896/jsjkx.231000189

Abstract

PDF(2353KB) ( 212 )

References | Related Articles | Metrics

With the continuous integration of informatization and industrialization,there are more and more intersecting parts between information domain and physical domain in industrial control systems,and network attacks on traditional information systems will threaten the industrial control system network.Traditional industrial control systems only consider the risks of functional safety,ignoring the impact of information security risks on functional safety.This paper proposes an integrated risk mode-ling method for functional safety and information security of industrial control system named SSPN-RA based on improved petri net,which includes three steps:integrated risk identification,integrated risk analysis and integrated risk assessment.This paper firstly identifies and abstracts the functional safety data and information safety data in the industrial control system,and then analyzes the collaborative attack path of functional safety and information security by constructing the petri net model combined with Kill Chain in the risk analysis process.Subsequently it quantifies the functional safety and information security nodes in the petri net,and finally calculates the risk value through the possibility of safety events and various losses caused by these safety events,so as to complete the integrated risk assessment of the industrial control system.In this paper,the feasibility of the proposed method is verified under the open-source simulation of chemical tank industrial control system,and compared with fault tree ana-lysis and attack tree analysis.Experimental results show that the proposed method can quantitatively obtain the risk value of industrial control system,and also solve the problem of cyber-physical collaborative attack and security risk that cannot be identified by the analysis of functional safety and information security.

Collaborative Network and Metric Learning Based Label Noise Robust Federated LearningMethod

WU Fei, ZHANG Jiabin, YUE Xiaofan, JI Yimu, JING Xiaoyuan

Computer Science. 2024, 51 (10): 391-398. doi:10.11896/jsjkx.230900050

Abstract

PDF(2268KB) ( 244 )

References | Related Articles | Metrics

Currently,there is limited research on the problem of label noise in federated learning.The main approaches involve introducing a benchmark dataset on the server side to evaluate the client's model,controlling the aggregation weights and feature class centers of the clients.However,most methods still have room for improvement in distinguishing noisy clients or noisy samples.This paper proposes a label-noise robust federated learning method based on co-networks and metric learning.The method consists of the following three parts:1) Client mutual evaluation mechanism.Clients score each other's models,construct a rating matrix,and further transform it into an adjacency matrix to differentiate clean/noisy clients.2) Collaborative network module.By constructing two collaborative equivalent federated network models,the Jensen-Shannon divergence is used to distinguish clean samples from noisy samples for the training of collaborative networks.3) Federated-collaborative network triplet loss.A loss function is designed to constrain the output features of the collaborative networks for the same noisy samples.Experimental verification is conducted on the publicly available datasets CIFAR-10 and CIFAR-100,and the results demonstrate the superiority of the proposed method in accuracy.

Identification of Mobile Service Type of Encrypted Traffic Based on Fusion of Inception andSE-Attention

WANG Yijing, WANG Qingxian, DING Dazhao, YAN Tingju, CAO Yan

Computer Science. 2024, 51 (10): 399-407. doi:10.11896/jsjkx.230900103

Abstract

PDF(3248KB) ( 265 )

References | Related Articles | Metrics

Mobile devices usually access WLAN and rely on WiFi encryption protocol to encrypt data link layer traffic in the network to maintain communication security.However,existing encrypted traffic identification methods mainly analyze traffic loads at the network layer and above,and cannot effectively identify the mobile service category of link layer encrypted traffic.To address this problem,a mobile service identification method based on link layer traffic in WiFi encryption scenarios is proposed.By passively sniffing WiFi data frames and extracting the traffic-side channel features available in the link layer,the traffic data is converted into a 2D histogram matrix.The recognition model,SE-Inception,is proposed by integrating the Inception network and SE-Attention mechanism,aiming to better capture the details and global information in the distribution features of traffic data frames,and highlighting the attention to important features to improve the recognition accuracy.In this paper,real datasets are used for experimental validation,and the results show that the method can effectively recognize the mobile service category of link-layer encrypted traffic in WiFi encryption scenarios,with an average accuracy of up to 98.29%,which is a better performance compared with the existing recognition methods.

System Call Host Intrusion Detection Technology Based on Generative Adversarial Network

FAN Yi, HU Tao, YI Peng

Computer Science. 2024, 51 (10): 408-415. doi:10.11896/jsjkx.230700014

Abstract

PDF(2367KB) ( 242 )

References | Related Articles | Metrics

The system call information of a program is an important data for detecting host anomalies,but the number of anomalies is relatively small,which makes the collected system call data often have the problem of data imbalance.The lack of abnormal system call data makes the detection model unable to fully understand the abnormal behavior pattern of the program,which leads to low accuracy and high false positive rate of intrusion detection.To solve the above problems,a system call host intrusion detection method based on generative adversarial network is proposed.By enhancing abnormal system call data,the problem of data imbalance is alleviated.Firstly,the system call trace of the program is divided into fixed length N-Gram sequences.Secondly,SeqGAN is used to generate synthetic N-Gram sequences from the N-Gram sequences of abnormal data.The generated abnormal data is combined with the original dataset to train the intrusion detection model.Experiments are carried out on a host system call dataset ADFA-LD and an Android system call dataset Drebin.The detection accuracy rate is 0.986 and 0.989,and the false positive rates is 0.011 and 0,respectively.Compared with the existing intrusion detection research methods based on hybrid neural network model,WaveNet,Relaxed-SVM and RNN-VED,the detection performance of the proposed method is better than other methods.

Attribute-based Sanitizable and Collaborative Data Sharing Scheme in Medical Scenarios

WANG Zheng, WANG Jingwei, YIN Xinchun

Computer Science. 2024, 51 (10): 416-424. doi:10.11896/jsjkx.230700187

Abstract

PDF(2277KB) ( 224 )

References | Related Articles | Metrics

Ciphertext policy attribute-based encryption(CP-ABE) is a mechanism that enables secure fine-grained access control of encrypted data,ensuring that only authorized users can access the data and avoiding unauthorized access and leakage in cloud environments to guarantee the safety of data.However,with the rapid development of cloud computing and IoT technology,traditional CP-ABE scheme is gradually unable to meet requirements of data sharing in terms of access policy expression and ciphertext security requirements in new medical IoT applications,such as multidisciplinary consultation,patient privacy data storage.This paper proposes an attribute-based sanitizable and collaborative sharing scheme in medical scenarios,which can effectively deal with malicious data owners by sanitizing ciphertext.Additionally,this method can specify collaborative nodes in the access structure,allowing users with different attribute sets to collaborate to obtain access rights.Security analysis shows that the proposed scheme has indistinguishable security under chosen plaintext attack.Performance analysis shows that compared with other schemes,the proposed scheme has lower computational overhead.

Black-box Adversarial Attack Methods on Modulation Recognition Neural Networks Based onSignal Proximal Linear Combination

GUO Yuqi, LI Dongyang, YAN Bin, WANG Linyuan

Computer Science. 2024, 51 (10): 425-431. doi:10.11896/jsjkx.230900054

Abstract

PDF(3258KB) ( 260 )

References | Related Articles | Metrics

With the extensive application of deep learning in the field of wireless communication,especially in signal modulation recognition,the vulnerability of neural networks to adversarial example attacks poses challenges to the security of wireless communication.Addressing the black-box attack scenario in wireless signals,where real-time feedback from the neural network is hard to obtain and only recognition results can be accessed,a black-box query adversarial attack method based on proximal linear combination is proposed.Initially,on a subset of the dataset,each original signal undergoes a proximal linear combination with target signals,where they are linearly combined within a range very close to the original signal(with weighting coefficients no greater than 0.05) and then input into the neural network to query.By counting the number of misrecognitions by the network for all proximal linear combinations,specific target signals most susceptible to linear combination effects for each original signal category are determined,which is termed the optimal perturbation signals.During attack testing,adversarial examples are generated by executing proximal linear combinations using the optimal perturbation signal corresponding to the signal category.Experimental results demonstrate that using the optimal perturbation signal for each modulation category on the chosen subset,the re-cognition accuracy of the neural network dropped from 94% to 50% when applied to the entire dataset,with a lower perturbation power compared to adding random noise attacks.Furthermore,the generated adversarial examples exhibit some transferability to structurally similar neural networks.This method,which generates new adversarial examples after statistical queries,is easy to implement and eliminates the need for further black-box queries.

Tripartite Evolutionary Game Analysis of Blockchain Applications in Patent Transactions

KANG Zhenwei, LI Jing, ZHU Jianming

Computer Science. 2024, 51 (10): 432-441. doi:10.11896/jsjkx.230800116

Abstract

PDF(4777KB) ( 249 )

References | Related Articles | Metrics

Blockchain technology helps to deal with problems such as opaque information and difficulty in confirming rights in patent transaction,and provides a new choice for patent transaction model innovation.By analyzing the role of blockchain in patent transaction,a tripartite evolutionary game model consisting of transaction applicant,transaction platform and intended transferee is constructed.Based on the game income matrix,the game evolution stability of patent trading system is analyzed,and the beha-vior strategy of the subject is discussed.On this basis,parameter values are assigned to different situations,and the influence of key factors on the patent trading system and the behavior strategy of each participant is verified by simulation analysis.The results show that the numerical relationship of parameters affects the evolutionary stability strategy of patent trading system.The blockchain application strategy of the trading platform is mainly affected by operating costs and losses caused by fraud.The increase of the concealing cost of under the role of blockchain can inhibit the fraudulent behavior of the transaction applicant.The reduction of investigation costs and losses can help to encourage the prospective transferee to choose the transaction.While enri-ching relevant theoretical research,it also provides reference for the application practice of blockchain in patent transactions.