Computer Science

Volume 53 Issue 5, 15 May 2026

Intelligent Education Technology

Personalized Learning Resource Recommendation:Classifications,Algorithms,and Challenges

SUN Yifei, LI Yongan

Computer Science. 2026, 53 (5): 1-12. doi:10.11896/jsjkx.250600184

Abstract ( 378 )

PDF(1890KB) ( 438 )

References | Related Articles | Metrics

Personalized learning resource recommendation represents an advanced form of deep integration between technology and education,which has received widespread attention from academia in recent years.Existing reviews have provided valuable discussions on basic theoretical frameworks,recommendation methods,and evaluation metrics.Building upon this foundation,the classification system and algorithmic analysis of personalized learning resource recommendations can be further enriched,especially through a comprehensive review of traditional recommendation methods and research on the application of emerging large language models in personalized learning recommendations.This review constructs a multi-dimensional classification framework,systematically categorizing personalized learning resource recommendations from three dimensions:application scenarios,algorithm categories,and learner characteristic modeling.It comprehensively explains the technical principles and latest developments of traditional recommendation algorithms,knowledge-based algorithms,machine learning algorithms,intelligent optimization algorithms,deep learning algorithms,and reinforcement learning algorithms.The review systematically analyzes the technical challenges facing personalized learning resource recommendations and the dilemmas from the learners’ perspective,and accordingly proposes future research directions such as integrating multiple recommendation methods,introducing educational theory guidance,and improving data quality.This review aims to provide educational technology researchers and practitioners with a systematic theoretical framework and technical roadmap to promote the continuous optimization and innovative development of personalized learning ecosystems.

Survey of Learning Trajectories

WANG Bixuan, CHEN Shiming, GAO Zhizezhang, FENG Jun, WANG Huiya

Computer Science. 2026, 53 (5): 13-21. doi:10.11896/jsjkx.250600159

Abstract ( 147 )

PDF(1667KB) ( 257 )

References | Related Articles | Metrics

With the development of intelligent education,learning trajectories have become a research hotspot.Its research aims to understand the development paths of learners,explore influencing factors,and provide a basis for educational decision-making.In recent years,the research on learning trajectories has achieved remarkable results in many aspects,such as the innovation of tra-jectory construction methods,the optimization of analysis methods,and the gradual expansion of application scopes.However,it cannot be ignored that this field still faces a series of severe challenges,including the fact that the theoretical system has not yet reached a unified standard,the integration degree among analytical methods is not high,and empirical analysis is relatively scarce.Meanwhile,at present,there is still a lack of comprehensive and systematic review studies on the learning trajectory.Based on the current development status of learning trajectory research,this article starts from two aspects:theoretical basis and technical practice,sorts out the main challenges faced in the current research,and takes the construction process,analysis methods and ty-pical application scenarios of learning trajectories as the clues to systematically summarize the relevant research results.On this basis,the development direction of future learning trajectory research is analyzed and prospected from two aspects:theoretical deepening and technological innovation,as well as application expansion.Proposing feasible optimization directions for the future is expected to promote the in-depth integration and application of learning trajectory analysis in the field of intelligent education.

Intelligent Analysis Technology on Teachers’ Teaching Emotions

LIU Yipu, MA Miao, HU Ximing

Computer Science. 2026, 53 (5): 22-29. doi:10.11896/jsjkx.250600163

Abstract ( 107 )

PDF(2178KB) ( 238 )

References | Related Articles | Metrics

Teachers’ positive teaching emotions can significantly improve teaching effectiveness,stimulate students’ interest and participation in learning,create a good classroom atmosphere,and promote students’ understanding and mastery of knowledge.On the contrary,inappropriate teaching emotions will affect students’ enthusiasm for learning,reduce teaching efficiency,and even lead to students’ anorexia.Under the background of the digital transformation of education in China,how to use the massive text,voice,video and other classroom teaching data for the intelligent detection and analysis of teachers’ teaching emotions is of great practical significance and value for improving the quality of teaching,enhancing the effects of teachers’ teaching reflections,and perfecting the teachers’ digital images.On the basis of traditional emotion analysis,this paper reviews the intelligent analysis technology of teachers’ teaching emotions based on artificial intelligence,including combing the research methods of teachers’ emo ions at home and abroad,analyzing the current research status of intelligent emotion analysis technology from the perspectives of text,speech,vision and multimodal,focusing on summarizing the key technologies of multimodal intelligent emotion ana-lysis technology,and finally focusing on analyzing the opportunities,challenges and development trends of the intelligent analysis technology of teachers’ teaching emotions.Finally,it focuses on analyzing the opportunities,challenges and development trends of the intelligent analysis technology for teachers’ teaching.

Application Advantages,Cases and Practical Challenges of Multimodal Technology in the Field of Education

LI Mengge, WANG Gang, BAI Wenhao, LEI Xue

Computer Science. 2026, 53 (5): 30-40. doi:10.11896/jsjkx.250600132

Abstract ( 412 )

PDF(3611KB) ( 295 )

References | Related Articles | Metrics

Education is the cornerstone of national development and national rejuvenation.However,the traditional education model has many limitations in teaching methods,resource allocation and evaluation systems,such as monotonous teaching me-thods,unbalanced educational resources and one-sided evaluation methods.With the rapid development of artificial intelligence technology,multimodal technology,as an emerging technology that integrates multiple data forms(such as images,sounds,and texts),provides new possibilities for solving these problems.Through applications such as smart classrooms and personalized learning systems,multimodal technology can comprehensively perceive and understand the learning environment,thereby breaking the limitations of the traditional education model,enhancing the learning experience,promoting educational equity and achieving personalized learning evaluation.Firstly,this paper outlines the definition,connotation and core algorithms of multimodal techno-logy,and explores its development trajectory and important position in the field of artificial intelligence.Secondly,this paper analyzes the application advantages of multimodal technology in the field of education in detail from multiple dimensions,and conducts in-depth discussions based on specific cases.Finally,this paper discusses the challenges faced by multimodal technology in educational applications,such as data privacy,technical costs and ethical issues.Through in-depth research on the applications and challenges of multimodal technology in education,this paper aims to provide theoretical basis and practical guidance for educatio-nal innovation,and promote the development of education towards a more intelligent,personalized and equitable direction.

Intelligent Analysis and Understanding Research Progress on Teachers’ Classroom Behavior

PAN Weiying, LI Yutong, MA Miao

Computer Science. 2026, 53 (5): 41-49. doi:10.11896/jsjkx.250600186

Abstract ( 317 )

PDF(2317KB) ( 198 )

References | Related Articles | Metrics

Against the backdrop of digital transformation in China,intelligent technologies represented by artificial intelligence,big data,the Internet of Things,and cloud computing have injected new impetus into education reform.Teachers’ classroom behavior is one of the important external manifestations of teachers’ professional qualities and teaching abilities.With the help of intelligent technologies such as integrated video analysis,speech processing,and text analysis,the characteristics and laws of teaching behaviors can be automatically represented,which is an important indicator for constructing digital portraits of teachers.This paper reviews the research progress of intelligent analysis and understanding of teachers’ classroom behavior.Firstly,an indicator system for intelligent analysis is constructed from three dimensions:teachers’ verbal behavior,non-verbal behavior,and the combination of verbal and non-verbal behavior.Then,the intelligent technologies,representative methods,and application practices in the recognition and understanding of teachers’ classroom verbal and non-verbal behaviors are summarized and sorted out from the perspectives of single modality and cross-modality.Finally,the current research challenges in the detection of tea-chers’ verbal behavior events,the semantic alignment of verbal and non-verbal behaviors,and the cross-modal information collaboration are discussed,as well as the future trends in the modeling of teachers’ verbal behavior events,semantic cross-modal fusion,and multi-modal large model application,which have important research significance and practical value for the evaluation of educational and teaching abilities and career development planning in the construction of teachers’ digital portraits.

Building 3L-S³ Smart Ecosystem:Systemic Transformation of Graduate Education Models andOrganizational Forms in the Age of Artificial Intelligence

DENG Xiaoheng, YU Zhan, XU Xuemei, LI Heng, ZHANG Hao, HU Chao

Computer Science. 2026, 53 (5): 50-58. doi:10.11896/jsjkx.250600135

Abstract ( 385 )

PDF(2039KB) ( 183 )

References | Related Articles | Metrics

Against the backdrop of a new generation of artificial intelligence permeating higher education,China’s graduate education is moving from “scale expansion” toward an “intelligent ecosystem”.Drawing on interim findings from the National Education Sciences Planning Project Graduate Education Models and Organizational Forms in the Age of Artificial Intelligence,this study advances the 3L-S³intelligent education model－an integrated framework of Learning-Lab-Launch(3L) and Smart Campus-Smart Governance-Smart Ethics(S³)－and explores how AI reshapes curricula,research training,organizational structures,and digital governance.A mixed methods design is adopted,combining grounded theory,a large scale survey of 1 862 respondents from the Graduate School of Central South University,an eight nation comparative analysis,and in depth case studies.The model’seffectiveness is further validated with three years of longitudinal data.The results indicate that:1)AI driven competence profiling increases personalized learning gains by 23.6%;2)A digital twin Meta-Lab shortens research cycles by 31.4%;3)Flattened,networked organizational forms have become this key carriers for talent-technology-governance synergy;4)Blockchain-AI collaborative governance raises the accuracy of academic misconduct early warning to 92%.In response,this paper proposes three policy pathways:1)Establishing a National AI Research and Education Center;2)Piloting interdisciplinary conversion master’s programs;3)Creating a Trusted AI Curriculum Consortium.Together,these contributions offer a replicable and assessable new paradigm for driving high quality advancement in China’s graduate education and for modernizing educational governance.

Research on Voice Cloning System Based on XTTS Model

WANG Chencai, YANG Siyan, MIAO Qiguang

Computer Science. 2026, 53 (5): 59-67. doi:10.11896/jsjkx.250600187

Abstract ( 133 )

PDF(4218KB) ( 215 )

References | Related Articles | Metrics

With the continuous advancement of deep learning and speech synthesis technologies,voice cloning has shown broad application prospects in intelligent voice assistants,virtual anchors,and barrier-free communication.However,existing voice cloning systems still face challenges in timbre similarity,interactive efficiency,and large-scale processing capability,making it difficult to meet the growing demand for high-quality,personalized speech synthesis.To address these limitations,this paper designs and implements a Web-based platform for multilingual voice cloning and batch text-to-speech synthesis,based on the XTTS mo-del.The system improves upon existing solutions by enhancing language coverage,reducing data dependency for timbre transfer,and optimizing batch processing efficiency.It adopts a front-end/back-end decoupled architecture,with a Flask-based RESTful API at the back end and mainstream Web technologies combined with AJAX at the front end.MySQL is used for managing user and audio data.The platform integrates voice cloning,text-to-speech,and batch synthesis modules,and demonstrates strong flexibility and scalability.Experimental results show that the system performs well in speech naturalness and timbre similarity,proving its practical value and application potential.

Learning Path Recommendation Based on Fusion of Hypergraph Neural Network and Dynamic Knowledge Tracking

LIU Meilin, MA Le

Computer Science. 2026, 53 (5): 68-78. doi:10.11896/jsjkx.250600157

Abstract ( 141 )

PDF(2116KB) ( 240 )

References | Related Articles | Metrics

Aiming at the deficiencies of the existing learning path recommendation methods in personalized adaptability,dynamic adaptability and multi-objective optimization,a collaborative recommendation model based on hypergraph neural network and knowledge tracking is proposed.By constructing an undirected graph of learning resources to encode the association relationship of resources,generating the embedding vectors of learning resources,and combining with the hypergraph neural network to aggregate the historical behavior data of learners,the interaction features between learners and learning resources are captured.It designs a multi-objective optimization strategy,utilizes the non-dominated sorting genetic algorithm(NSGA-II) to generate the Pareto frontier solution set,synchronously optimizes the accuracy,difficulty adaptability,effectiveness and diversity of learning resources of the learning path,and combines weight distribution and comprehensive utility function to improve the quality of the learning path.Experiments show that the proposed method is tested on the MOOCCube and MOOPer datasets.Among them,the HR@5 and MRR@5 of the MOOPer data reach 93.9% and 90.7% respectively,achieving precise recommendation of learning paths and verifying the effectiveness of the model in the modeling of learners’ historical interactions and the integration of course structure constraints.

Temporal Dynamic Tag Construction and Prediction Method for Learner Profile

BAI Jinghao, ZHUANG Junxi, LAI Yingxu

Computer Science. 2026, 53 (5): 79-89. doi:10.11896/jsjkx.250400012

Abstract ( 99 )

PDF(4145KB) ( 251 )

References | Related Articles | Metrics

With the advancement of educational informatization and the growing emphasis on personalized learning,how to construct accurate and dynamic learner profiles from massive educational data has become a crucial research topic in the field of education.However,existing approaches to learner profiling often suffer from insufficient feature representation,inadequate conside-ration of temporal dynamics,and inconsistencies in tag dependencies.To address these challenges,this paper proposes a method for constructing and predicting temporally dynamic tags.Specifically,the proposed method enhances original feature tags through a multi-source learner feature augmentation strategy,and performs temporal tag modeling under different observation periods based on educational theories.Furthermore,it leverages Teacher Forcing and Scheduled Sampling techniques from natural language processing(NLP) to train a multi-branch Transformer model that deeply captures learner features across varying temporal scales.The proposed approach is evaluated on two publicly available datasets－campus card swipe data and online course records,and the results demonstrate its effectiveness in constructing dynamic temporal tags and learner profiles.

Innovative Automated Scoring Based on Large Language Models

WANG Shenghui, LI Teng

Computer Science. 2026, 53 (5): 90-98. doi:10.11896/jsjkx.250600183

Abstract ( 208 )

PDF(2416KB) ( 249 )

References | Related Articles | Metrics

Innovative automated scoring(IAS) is crucial in education.Traditional scoring is subjective,inefficient,and lacks uniform standards.The fast progress of large language models offers new solutions.This study creates a high-quality dataset WAIS and presents a semantic-driven hierarchical topic extraction algorithm.Through four phases-semantic chunking,basic topic extraction,optimized analysis,and topic fusion-the algorithm improves the model’s ability to extract themes from student answers,enabling automatic topic extraction.It offers a solid basis for automated scoring and establishes an explainable cognitive framework for subsequent scoring.The study compares three prompting strategies:Zero-shot,Few-shot,and Chain-of-Thought(CoT),and evaluates them using several pre-trained models.Results show CoT is superior.The DeepSeek-R1 model achieves 68% accuracy.After fine-tuning,the smaller-parameter model Qwen1.5-7B reaches 83% accuracy,even slightly surpassing the larger-parameter model using only the prompt in innovative scoring tasks.This indicates that using large language models for innovative automated scoring is feasible and has great potential for development.

Multimodal Continuous Emotion Recognition for English Spoken Emotion Evaluation

WANG Liyan, ZHANG Qian, GUO Yuanyuan, CHEN Haifeng, LI Jian

Computer Science. 2026, 53 (5): 99-108. doi:10.11896/jsjkx.250600162

Abstract ( 93 )

PDF(3426KB) ( 205 )

References | Related Articles | Metrics

Spoken English occupies a crucial position in English learning.Addressing the scarcity of existing datasets for evaluating emotional expression in spoken English and the inadequate utilization of multimodal information,this paper introduces a novel dataset named the English spoken multimodal emotion dataset(ESMED).This dataset is annotated with continuous emotions(arousal,valence) and emotional quality scores.Additionally,an innovative network model for evaluating spoken English emotions is proposed.The model initially compresses and fuses continuous emotional information through perception resampling and multimodal fusion modules to predict arousal and valence.Subsequently,it performs specific transformations on the features through learnable bottleneck and joint decoding layers.The emotional quality evaluation module then jointly decodes arousal,valence,and transformed features to obtain the final quantified emotional quality score.Experimental results demonstrate that the proposed model achieves a concordance correlation coefficient(CCC) of 0.500 3 and a mean absolute error(MAE) of 0.635 4 on the ESMED dataset,verifying the effectiveness and accuracy of the proposed method.

Database & Big Data & Data Science

Review of Uniform Manifold Approximation and Projection

ZHANG Run, LI Xiaobin, XU Yamin

Computer Science. 2026, 53 (5): 109-118. doi:10.11896/jsjkx.250400084

Abstract ( 143 )

PDF(3019KB) ( 203 )

References | Related Articles | Metrics

This paper systematically introduces the theoretical foundations,algorithmic implementations,and recent developments of UMAP algorithm.UMAP is a nonlinear dimensionality reduction method grounded in algebraic topology and category theory.It aims to preserve the local structure of high-dimensional data while faithfully capturing its global distribution.The algorithm can be divided into two stages.Firstly,it constructs a weighted k-nearest neighbor graph that encodes local geometric relationships,with neighborhood sizes adaptively determined by data density.Secondly,it learns a low-dimensional embedding by minimizing the cross-entropy between the high-dimensional and low-dimensional neighborhood graphs,thereby preserving topological structure and enabling effective visualization.This survey further reviews several notable extensions of UMAP.1) Supervised and Semi-supervised UMAP enhances class separation by incorporating label information;2) Parameterized UMAP integrates neural networks to achieve generalizable nonlinear mappings;3) DensMAP preserves data distribution characteristics through density correlation optimization;4) AlignedUMAP enables aligned embeddings across datasets or temporal sequences;5) Progressive UMAP addresses dynamic embedding challenges for streaming data and out-of-sample extensions;6)GNUMAP,which combines with graph neural networks;7)MultiMAP,which is applied to multimodal data fusion.

Data Resource Organization Method Based on Enterprise Dataspace and Data Asset Management

LI Minbo, WANG Shaohua, WU Dazhen

Computer Science. 2026, 53 (5): 119-128. doi:10.11896/jsjkx.250600019

Abstract ( 142 )

PDF(2868KB) ( 198 )

References | Related Articles | Metrics

Aiming at the problem of data resource islands caused by confidentiality hierarchical and block management in military research institute,and the difficulty of intelligent retrieval and knowledge reuse of data resources,a governance solution for multi-source heterogeneous enterprise data resources and data assets is proposed.The graph node association mapping between data resources is realized through the attribute graph model of enterprise data space,and a data resource knowledge graph integrating the BOM tree structure is constructed,covering the hierarchical relationship,attribute information and association relationship of R&D process,production and manufacturing,and quality inspection data.This paper proposes a novel RAG framework－HireRAG,and establishes a community-based hierarchical index of knowledge graph based on C-HNSW.The low-level retains fine-grained knowledge units,and the high-level community provides a global summary to handle retrieval at different levels.A graph-enhanced clustering algorithm is proposed to enable C-HNSW to better capture the semantic information in the know-ledge graph.Experiments demonstrate that HireRAG is more adapt at processing bill of materials(BOM) related data within enterprise data spaces compared to several existing advanced retrieval-augmented generation(RAG) frameworks.Furthermore,it achieves superior performance metrics in both retrieval recall and accuracy.The data asset management system ensures that the data assets are entered into the table in compliance with the whole process.

Deep Learning Training Time Prediction Algorithm Integrating Multi-dimensional Operator Features

CHEN Yuansheng, CHEN Shunjue, MO Xuan, WU Weigang, LI Jialun

Computer Science. 2026, 53 (5): 129-136. doi:10.11896/jsjkx.250900001

Abstract ( 128 )

PDF(2215KB) ( 219 )

References | Related Articles | Metrics

Offline tasks are delay-tolerant workloads without strict requirements on completion time,typically including batch processing or machine learning tasks.With the development of deep learning technology,deep learning tasks have become one of the important parts of offline workloads in cloud data centers.Accurate runtime prediction of offline tasks improves resource utilization during idle periods of online tasks.However,deep learning models exhibit diverse architectures and vast scale differences.Factors such as batch sizes,hyperparameters and operator characteristics during training also significantly affect task execution time.Existing methods struggle to comprehensively account for all these factors:configuration-based methods ignore the internal execution mechanism of the algorithm;operator-based methods neglect the impact of computation graph structure;graph-based methods either face excessive model complexity with graph neural networks or lose dependency information when simplifying to topological sequences.In view of the deficiencies of the topological sequence methods,this paper proposes the MDOT(Multi-dimensional Operator Transformer) algorithm to convert the computational graph into an operator sequence according to topological sorting.Based on this sequence of operators,MDOT uses Transformer to fuse the three-dimensional information of the operators:operator type,operator configuration,and computational load to perform multi-dimensional operator encoding,more comprehensively modeling the execution characteristics of the operators.Secondly,in order to capture the dependencies of the computational graph,MDOT designs a graph position encoding mechanism,which captures the relationships between operator sequences through the self-attention of the Transformer and models the mutual influence of operators in terms of running time.Experimental results show that MDOT outperforms existing methods in predicting the training time of deep learning tasks,with the mean absolute error and root mean square error being 25% and 45% lower than those of suboptimal models,respectively.

Efficient Algorithm for Counting the Shortest Cycle on Temporal Graphs

LI Tianqi, DU Ming, ZHOU Junfeng

Computer Science. 2026, 53 (5): 137-148. doi:10.11896/jsjkx.250400006

Abstract ( 176 )

PDF(3472KB) ( 205 )

References | Related Articles | Metrics

Shortest cycle counting serves as a fundamental approach for analyzing graph data,with critical applications in key node identification,periodic structure analysis,and anomaly detection.However,in temporal graph analysis,existing methods face significant challenges in efficiently supporting shortest cycle counting across varying time windows due to the temporal nature of edges and query window requirements.For this issue,an index-based solution is proposed to enhance counting efficiency firstly,which is followed by an optimized index construction strategy that reduces index size while improving construction efficiency,so as to ensure both compact index storage and efficient query processing without compromising performance.Finally,experimental results on real-world temporal graphs demonstrate that the proposed method can efficiently construct the index and rapidly retrieve the count and length of shortest cycles passing through the given query vertice within arbitrary time windows.

Optimized Algorithm for Colorful h-star Core Decomposition

LIANG Yue, ZHOU Junfeng, DU Ming

Computer Science. 2026, 53 (5): 149-156. doi:10.11896/jsjkx.250300166

Abstract ( 158 )

PDF(3074KB) ( 151 )

References | Related Articles | Metrics

The colorful h-star core is a cohesive subgraph model defined based on colorful h-stars,which has a wide range of applications in understanding the complex relationships of networks and analyzing the higher-order structure of communities.However,the existing colorful h-star core decomposition algorithms have significant limitations:as the h-value increases,the h-star degree grows exponentially,which is prone to data overflow and leads to erroneous computation results.To address this problem,this paper proposes a pruning method to identify the vertices that have no effect on the h-star core number,and further derives a tight upper bound for the h-star core number by mathematical derivation,which can effectively replace the h-star degree during the core decomposition process,thus solving the problem of computational errors due to the excessive h-star.Experimental results show that the proposed method can correctly compute the h-star core decomposition with more than 10 times improvement in computational efficiency.

Multi-scale Transformer Oil Price Prediction Framework with AEMD and Trend Cross-attention

LI Tengjia, MA Chun’ai

Computer Science. 2026, 53 (5): 157-163. doi:10.11896/jsjkx.250900086

Abstract ( 104 )

PDF(3119KB) ( 208 )

References | Related Articles | Metrics

Against the backdrop of global low-carbon transition and energy restructuring,crude oil price forecasting has become not only a key topic in energy market analysis but also an essential reference for policy-making and investment decisions.How-ever,crude oil price series often exhibit strong nonlinearity and pronounced non-stationarity.Existing methods still face limitations in feature extraction and temporal modeling:on the one hand,the mining of multi-scale features is often insufficient,leading to biased characterization of short-term fluctuations and long-term trend evolution;on the other hand,the integration of short- and long-term information is frequently mishandled,making it difficult to balance predictive accuracy with trend stability.To address these challenges,this paper proposes a multi-frequency decoupled dual-branch Transformer model(MFD-DBV-Transformer) for Brent crude oil price forecasting.The method firstly employs adaptive empirical mode decomposition(AEMD) to decompose the crude oil price series into multiple intrinsic mode functions(IMFs).This distinguishes high-frequency short-term components from low-frequency long-term trends.An adaptive frequency decoupling module(AFDM) is then designed to construct dual-branch feature representations,separately capturing short-term volatility patterns and long-term trend features.A trend fusion module is further introduced,where cross-attention is used to achieve adaptive modulation of short-term predictions with long-term trend information.In addition,a temperature-regulated adaptive masking mechanism is incorporated to prevent overfitting in long-term trend modeling and to enhance the model’s generalization ability in volatile market environments.Experimental results demonstrate that the proposed MFD-DBV-Transformer achieves superior performance in capturing complex time-frequency characteristics of crude oil prices,significantly outperforming traditional LSTM and several mainstream deep learning models.The model not only improves forecasting accuracy but also demonstrates stronger stability and adaptability in trend tracking.The proposed approach provides policymakers and energy investors with an efficient and reliable forecasting and decision-support tool for coping with crude oil market volatility,while offering new insights and methodologies for modeling complex non-stationary time series.

Structure-aware Trace Clustering Method Based on Process Edit Distance

YE Jianhong, WU Yongjin, HUANG Hongkai

Computer Science. 2026, 53 (5): 164-173. doi:10.11896/jsjkx.260100070

Abstract ( 111 )

PDF(2846KB) ( 212 )

References | Related Articles | Metrics

In model checking and process mining,trace clustering groups similar execution traces to support the construction of accurate behavioral models,the verification of model correctness,and data-driven model refinement.However,existing sequence pattern-based trace clustering approaches typically treat traces as ordinary strings,neglecting the inherent concurrent and cyclic execution relationships among activities.This simplification often leads to the loss of structural information and consequently degrades clustering quality.To address this issue,this paper proposes a novel trace similarity measurement method,referred to as process edit distance.The proposed method first normalizes concurrent execution relationships among activities into a consistent sequential representation.It then abstracts repetitive loop behaviors through a compression and simplification mechanism to reduce the influence of redundant executions.Finally,trace similarity is measured by jointly considering activity occurrences and the direct-follow relationships between activities.Furthermore,to obtain process models that better reflect real business behavior,a post-processing strategy termed merging noise clusters is introduced within an agglomerative hierarchical clustering framework to alleviate structural fragmentation caused by noise or small-sized clusters.Experimental results demonstrate that the trace clustering algorithm based on process edit distance outperforms existing methods of the same category in terms of clustering quality,while exhibiting strong stability and robustness.In addition,the merging noise clusters strategy consistently reduces the overall structural complexity of the clustering results,leading to clearer and more interpretable process models.

Computer Graphics & Multimedia

Fake News Video Detection:Methods,Challenges,and Explainability Research

LI Yili, YAO Jietong, LANG Jian, ZHU Guobin, CHEN Leiting, ZHOU Fan

Computer Science. 2026, 53 (5): 174-192. doi:10.11896/jsjkx.250900048

Abstract ( 92 )

PDF(3302KB) ( 185 )

References | Related Articles | Metrics

As video platforms have become major carriers of news dissemination,their unique technological characteristics have facilitated the wide spread and rapid diffusion of fake news.The proliferation of video fake news has already caused significant harm in domains such as politics,public health,and the economy.Existing surveys have mostly focused on text-based or image-based fake news detection,with a lack of systematic overviews specifically addressing video fake news detection.To fill this gap,this survey presents,for the first time,a comprehensive synthesis of existing detection methods and available datasets for video fake news.On this basis,the survey also highlights its novelty by introducing explainability as an independent dimension of analysis.This additional perspective explicitly addresses the practical needs for auditability,interpretability,and traceability in real-world detection scenarios.Specifically,in order to construct a clear conceptual framework,this survey first provides a formal definition of what constitutes video fake news.Building upon this foundation,the survey proceeds to categorize the wide range of existing detection methods into three major groups,namely intrinsic -feature -based methods,external-cue -based methods,and explainable methods.The intrinsic -feature -based category emphasizes the direct analysis of the multimodal content contained within the videos themselves.By contrast,external-cue -based methods shift attention beyond the video content and make use of auxiliary signals such as patterns of user behaviors,structures of information propagation across social networks,and platform-level metadata in order to provide supportive verification.Finally,explainable methods are distinguished by their focus on transparency and interpretability.Rather than offering only a binary classification label,these approaches are designed to generate explicit decision rationales.Subsequently,this survey summarizes existing video fake news datasets,including user-generated content,professionally produced media,and explainability-enhanced resources.Finally,it discusses the key challenges and limitations of video fake news detection and outlines promising directions for future research.

Anomaly Detection and Localization Technology for Gravity Wave Spectral Images Based onPre-trained Networks

HUANG Siyang, YAO Ye, ZHU Yian, HAI Duo, XIONG Zhihai

Computer Science. 2026, 53 (5): 193-206. doi:10.11896/jsjkx.250400117

Abstract ( 103 )

PDF(4568KB) ( 207 )

References | Related Articles | Metrics

To addresses the issues of indistinct features and uneven distribution in gravitational wave spectral image data,which often lead to high error rates in anomaly detection.This paper proposes an anomaly detection method for gravitational wave spectral images based on a pre-trained network.This method analyzes image features at both the image level and pixel level,employing preprocessing techniques to enhance the key features of the images,thereby more accurately capturing useful feature information within the images.The intermediate layers of an ImageNet pre-trained network are utilized for feature extraction,and a core-set subsampling mechanism is applied to compress the feature memory bank,reducing inference analysis time.Finally,the nearest neighbor algorithm is used to calculate the anomaly scores of image pixels,enabling the assessment of the overall anomaly degree of the image and the identification of anomalous regions.Experimental results demonstrate that this method can effectively analyze features of gravitational wave spectral images at both image and pixel levels,utilize image features for anomaly detection,and accurately identify anomalous regions in gravitational wave spectral images.The AUROC metrics for anomaly discrimination and localization reach 98.73% and 95.19%.

uHairDet:Method for Synthetic Hair Image Generation and Detection in Androgenetic AlopeciaDiagnosis

CHEN Qi, CHEN Xingkai, ZHANG Huihuang, SU Yiping, HU Haigen

Computer Science. 2026, 53 (5): 207-217. doi:10.11896/jsjkx.251100057

Abstract ( 203 )

PDF(4795KB) ( 217 )

References | Related Articles | Metrics

The clinical diagnosis of Androgenetic Alopecia(AGA) relies heavily on manual hair counting,a process that is time-consuming,subjective,and limits diagnostic efficiency.While automated hair recognition technology holds great promise,it faces challenges due to fine hair structures and a lack of pixel-level annotated data.To address this,this paper proposes uHairDet,a novel framework that minimizes dependency on manual annotations by leveraging synthetically generated data.The proposed approach consists of three key components:1)a hair synthesis with built-in annotations data generator(HBDG) for creating structurally plausible hair images with pixel-level labels;2)a structure-stable style-transfer GAN(BS-GAN) incorporating a semantic-aware adaptive error correction loss(HEE Loss) to enhance structural consistency;3)an FCOS+PSC detection model based on a Mean-Teacher framework,trained with a new oriented bounding box(OBB) annotation paradigm tailored for hairs to preserve critical information.Results demonstrate that the proposed method,requiring no manual labels,achieves a 56.9% AP,significantly outperforming baseline models and establishing a new paradigm for intelligent assisted diagnosis in hair-related disorders.

Frequency Driven Multi-scale Image Super-resolution Method

YANG Hongju, ZHANG Ziyang, LI Yao

Computer Science. 2026, 53 (5): 218-227. doi:10.11896/jsjkx.250700046

Abstract ( 85 )

PDF(5339KB) ( 212 )

References | Related Articles | Metrics

Single-image super-resolution is a critical technique aimed at reconstructing high-resolution images from single low-resolution inputs.It plays a pivotal role in various fields such as medical image enhancement,satellite remote sensing,security surveillance,and digital media content optimization.The core of SISR lies in restoring the lost high-frequency details to improve visual quality.However,this task faces numerous challenges due to its ill-posed inverse problem nature:traditional interpolation methods struggle to reconstruct complex details;convolutional neural network-based approaches can extract local features but fall short in global modeling;while Transformers excel at capturing long-range dependencies,they have limited capabilities for refining high-frequency information,leading to blurred edges and distorted textures in reconstructed images.To address these issues,this study proposes an innovative model that leverages wavelet transformation for multi-scale image decomposition,optimizing the extraction of high-frequency information.Three key modules are designed:the wavelet refinement module enhances the processing of high-frequency details;the shifted rectangular feature enhancement module captures global context;and the multi-scale wavelet fusion module integrates high-frequency priors with global features.This method significantly improves texture clarity and edge sharpness,balancing both local details and overall consistency.Experimental results demonstrate that the proposed model outperforms existing techniques on benchmark datasets such as Set5,Set14,and BSD100,achieving an average peak signal-to-noise ratio improvement of approximately 0.3 dB,along with superior subjective visual quality.This research not only effectively tackles the challenge of high-frequency information recovery but also provides new insights into the field of single-image super-resolution,holding significant academic and practical value.

High-accuracy Human Pose Estimation Combining Wavelet Analysis and Frequency-DomainAttention

LI Zongmin, WANG Li, LI Yachuan, LIU Yujie, RONG Guangcai, LIU Weihan, MA Wenkang

Computer Science. 2026, 53 (5): 228-236. doi:10.11896/jsjkx.250800025

Abstract ( 244 )

PDF(4120KB) ( 191 )

References | Related Articles | Metrics

HPE(Human Pose Estimation) is a fundamental task in computer vision,aiming to accurately localize human keypoints and understand body structure,which is crucial for downstream tasks such as action recognition and detection.Although deep learning has driven significant progress in HPE,existing methods still struggle to effectively handle challenges like large scale variations,occlusion,and loss of details in complex scenarios such as dense crowds and dynamic movements with large pose changes.To address these issues,this paper proposes an improved architecture,EFW-HRNet,which fuses DWT(Discrete Wavelet Transform) with the HRNet(High-Resolution Network).It introduces DWT-based downsampling and feature fusion modules to capture and preserve multi-scale details.It designs a CBA(Cross Band Attention) module to enable adaptive interaction among DWT sub-band features and enhance robustness against occlusion.And it applies a FBCC(Frequency Band Channel Compression) strategy to compress high-frequency channels,significantly reducing computational redundancy and improving model efficiency.Experiments on the COCO dataset show that EFW-HRNet achieves a significant AP increase of 4.0 percentage points compared to the strong baseline UDP HRNet-W32.Ablation studies validate the effectiveness of the DWT,CBA,and FBCC strategies,where FBCC achieves a good trade-off between accuracy and efficiency,sacrificing only about 0.8 percentage points AP in exchange for a substantial reduction in parameters by about 66% and computational cost by about 51%.

Continuous Image Super-resolution Based on Self-attention Implicit Feature Encoding andDecoding

CHEN Boying, SHI Jie

Computer Science. 2026, 53 (5): 237-246. doi:10.11896/jsjkx.250400097

Abstract ( 121 )

PDF(3470KB) ( 210 )

References | Related Articles | Metrics

In order to solve the problem that the convolutional neural network can only deal with fixed resolution in image super-resolution reconstruction,and it is difficult to realize continuous super-resolution reconstruction with arbitrary resolution,this paper proposes an algorithm for super-resolution reconstruction of continuous-rate images based on self-attention implicit feature encoding and decoding.Firstly,the feature encoding network based on local-global self-attention modeling is used to realize the feature mapping from low-resolution image to high-dimensional feature.Then local implicit feature enhancement encoders are used to effectively aggregate local neighborhood features,and self-attention mechanism is used to enhance the correlation between feature neighborhood data.Finally a multi-granularity implicit feature decoder is used to predict the pixel values of high-resolution coordinates by inputting the image coordinates and the depth features of the adjacent coordinate multilayer perceptron.Experimental results show that,compared with the current image reconstruction algorithm,the proposed method achieves better super-resolution reconstruction results,which proves the superiority and effectiveness of the method.

Artificial Intelligence

Memory Modeling Based on Dynamic Distributed Directed Graph

WEI Hui, FENG Chenyue

Computer Science. 2026, 53 (5): 247-256. doi:10.11896/jsjkx.250400093

Abstract ( 151 )

PDF(4964KB) ( 214 )

References | Related Articles | Metrics

Memory mechanism research is a core topic in computational neuroscience.While existing studies have revealed the importance of synaptic plasticity,signal transmission,and other factors in memory formation within neural networks,accurately replicating these complex biological mechanisms in computational models remains a significant challenge.Traditional graph-theoretical models can describe the topological structure of neural networks,but their static nature and centralized information processing cannot fully capture the dynamic,distributed,and decentralized characteristics of biological neural networks.Therefore,there is an urgent need for a memory model based on dynamic directed graphs that better aligns with the features of biological neural systems.The model is based on memory trace theory and employs a sparsely connected directed graph network structure,allowing autonomous decision-making through local information to achieve decentralized parallel processing.The model innovatively introduces a variable resistance structure,dynamically adjusting resistance values to simulate synaptic plasticity in neurons.Additionally,a resource-competition-based path reinforcement mechanism is used to mimic the memory process in biological neural networks.Experimental results show that the model consistently achieves memory functionality across various network scales and topologies,and the memory capacity increases approximately linearly with network width,exhibiting characteristics like those of biological neural networks.Compared to existing mainstream models,the proposed model demonstrates significant advantages in resource efficiency,memory capacity,and network scalability,providing a solid theoretical foundation for the design of brain-like computing systems and the development of intelligent systems.

Robust Incremental Fuzzy Concept-cognitive Emotion Recognition Method Based on Three-wayDecision

XU Weihua, HU Kaiping

Computer Science. 2026, 53 (5): 257-267. doi:10.11896/jsjkx.260300053

Abstract ( 106 )

PDF(2850KB) ( 209 )

References | Related Articles | Metrics

Speech emotion recognition(SER) plays an important role in human-computer interaction systems.In order to solve the problems that the decision making process of existing deep learning models is opaque in the SER task,and the traditional concept-cognitive learning(CCL) is susceptible to noise interference and concept drift when processing incremental data,a three-way weighted fuzzy concept-cognitive classification framework that leverages extremely randomized trees(3WERT-WFCCL) is proposed.In the feature processing,Whisper is used to extract high dimensional speech features,and a multi-layer perceptron(MLP) is used for hierarchical abstract representation.In the cognitive learning stage,the extremely randomized trees(ERT) algorithm is introduced to calculate the importance of features to realize the automated quantitative allocation of attribute weights,and the three-way decision fault tolerance threshold parameter is embedded in the cognitive operator to construct a positive and negative two-way cognitive mechanism.In the face of incremental data,the model divides the new samples into a positive region,a boundary region and a negative region according to the feature identification distance,and adopts a robust strategy that only uses the positive region samples to update the concept,which effectively resists the noise interference.On the SAVEE dataset with more complex feature boundaries,the robust update strategy improves the accuracy by 0.16 percentage points compared with the global update strategy.Experiments on two public datasets EmoDB and SAVEE show that 3WERT-WFCCL is superior to the existing baseline methods in multiple key evaluation indicators.Compared with the baseline models Logistic Regression(LR) with the best performance on each data set,the accuracy of the proposed algorithm is increased by 1.53 percentage points and 0.62 percentage points respectively,and the F1 score is increased by 1.28 percentage points and 0.40 percentage points respectively.Experimental results verify the effectiveness of the three-way decision mechanism,which provides a new method for constructing SER models with high classification accuracy,strong noise robustness and logical interpretability.

Construction of Chinese-Burmese Machine Translation Corpus Based on Pivot OptimizationSelf-training

LAI Hua, GUO Zirui,LI Ying, YU Zhengtao

Computer Science. 2026, 53 (5): 268-275. doi:10.11896/jsjkx.250300142

Abstract ( 143 )

PDF(2599KB) ( 157 )

References | Related Articles | Metrics

In recent years,the rapid development of language models has greatly promoted the model effect of supervised machine translation.However,the performance of supervised machine translation is highly dependent on the quality of parallel corpora.In view of the lack of high-quality Chinese-Burmese parallel corpora resources,this paper proposes a corpus construction method based on pivot optimization self-training.Firstly,the initial machine translation model is trained with a small-scale high-quality Chinese-Burmese parallel corpus.Then,a pseudo-parallel corpus from Burmese to Chinese is generated based on this model.At the same time,an English-Burmese parallel corpus with English as the pivot language is introduced,and the pivot English is translated into Chinese using existing high-quality English-Chinese translation tools to construct a second set of Chinese-Burmese pseudo-parallel corpora.To further improve the quality of the pseudo-parallel corpus,it designs a cross-lingual representation scoring mechanism to select higher quality sentence pairs from the two sets of pseudo-parallel corpora based on semantic similarity.Finally,the initial translation model is iteratively optimized and trained using the selected high-quality pseudo-parallel corpora.Experimental results show that the proposed method achieves an average 8.32 BLEU value improvement in the Chinese-Burmese machine translation task.Detailed analysis experiments prove that the pivot language optimization method can effectively enhance the model self-training effect and gradually improve the quality of pseudo-parallel corpus when the initial model performance is weak.In addition,this study constructs 700 000 high-quality Chinese-Burmese parallel corpus to further promote the development of Chinese-Burmese machine translation.

Boosting Generative Rule Extraction via Negative-aware Approach

JI Wendi, WANG Yongquan, SHEN Yicheng

Computer Science. 2026, 53 (5): 276-285. doi:10.11896/jsjkx.250400141

Abstract ( 122 )

PDF(4331KB) ( 192 )

References | Related Articles | Metrics

Legal rules are behavior norms formulated by competent authorities with binding legal force and effectiveness,essential to maintaining social order.As the preliminary of Legal AI,numerous studies have attempted to convert natural-language legal texts into machine-readable rule sets,however the results remain unsatisfactory.To address the challenges of formal representation and extraction of legal rules,this paper proposes a systematic legal rule extraction paradigm and introduces a negative-aware approach to boosting generative rule extraction with large language models(LLMs).The paradigm defines a legal rule schema by decomposing a legal rule into a tetrad of subject,object,conditions and consequences,thus clarifying its applicability,targets and effects.Building on this,this paper proposes a generative legal rule extraction enhancement method leveraging LLMs,which incorporates the concept of “learning from errors” by constructing a negative-aware training framework to improve the model’s ability to recognize hard negative cases and mitigate hallucination issues in generative rule extraction.Experimental results show that the rule extraction model based on Mistral-Small-24B(a mid-size LLM) outperforms the general-purpose LLM(Deepseek-r1) by 18.23% and even surpasses human-annotated performance by 1.5%,demonstrating that the negative-aware training framework significantly enhances the rule extraction capability of the model.

Explainable Sentencing Prediction Method Driven by Sentencing Rule Knowledge Graph

HAN Linrui, ZHENG Ri, CONG Yingnan

Computer Science. 2026, 53 (5): 286-298. doi:10.11896/jsjkx.251000076

Abstract ( 189 )

PDF(5088KB) ( 206 )

References | Related Articles | Metrics

Sentencing prediction stands as a core task for legal artificial intelligence empowering criminal justice,playing a vital role in overcoming sentencing bias,enhancing judicial efficiency,and safeguarding fairness and justice.Addressing the bottleneck issues of low prediction accuracy and insufficient interpretability inherent in traditional machine learning models,this paper proposes an explainable sentencing prediction method driven by sentencing rule knowledge graph.The method innovatively designs a knowledge graph and large language model integration architecture.The technical roadmap is as follows.Firstly,a structured sentencing rule knowledge graph is constructed top-down using the BERT-BiLSTM-CRF model.Subsequently,Chain-of-Thought reasoning for sentencing is distilled from the Sentencing Guidelines,and structured prompting based on the graph’s data is employed to conduct supervised instruction fine-tuning on large language models(LLaMA-3-8B-Chinese-Chat,Qwen-2-7B,Baichuan2-7B-Chat,GLM-4-9B-Chat),guiding them to learn standardized sentencing reasoning logic.Finally,during the prediction phase,retrieval-augmented generation is implemented on the fine-tuned model via the graph’s entity recognition and retrieval mechanism,outputting sentencing predictions alongside explainable step-by-step analyses consistent with sentencing rules.Expe-rimental results demonstrate that:1)BERT-BiLSTM-CRF model achieves an F1 score of 0.953 8 on the entity-relation extraction task,outperforming conventional models;2)GLM-4-9B-Chat model achieves the best overall performance in both test-set generation quality and downstream tasks;3)The final sentencing prediction model achieves an F1 score of 0.627 6,significantly outperforming baseline models such as MTL-Fusion,Lawformer,and BERT.Moreover,generating explanatory text following the standardized logic of “determining the sentencing starting point-baseline sentence-adjusting the baseline sentence-declared sentence” significantly enhances the interpretability and user acceptance of results;4)Ablation studies and human evaluations jointly demonstrate the model’s significant superiority over baselines in sentencing accuracy,precision of legal provisions citation,logical cohe-rence and fluency of reasoning,as well as compliance with standardized sentencing steps.This research establishes a novel paradigm integrating knowledge-driven and data-driven approaches for legal AI.

EC-MIIP:Efficient Fine-tuning Small-parameter Large Language Model for Intellectual Property

LIU Xukai, LIU Yang, HUANG Haozhen

Computer Science. 2026, 53 (5): 299-308. doi:10.11896/jsjkx.250600023

Abstract ( 89 )

PDF(3012KB) ( 188 )

References | Related Articles | Metrics

In recent years,large language models have been developing rapidly,demonstrating excellent capabilities in several na-tural language processing tasks,and providing strong technical support in the field of intelligent justice.Combining model pre-training and fine-tuning techniques,this paper constructs a database of MIPLD(Micro-model Intellectual Property Learning Direction) intellectual property directions under small parameters,and realizes an algorithmic framework for distributed pre-training according to the characteristics of the discipline of law and the characteristics of the intellectual property systems.Subsequently,based on the database of MIPLD,high-quality fine-tuned Q&A pairs of each direction are constructed,and EC-MIIP,an intellectual property problem analysis model with high capacity density under small parameters,is realized,which can be used for tasks such as intellectual property doctrine quizzing,analysis the nature of the act,judicial case analysis,and legal document writing.Experimental results show that EC-MIIP performs better than Owen3-4B,Qwen3 full-parameter and Deepseek-R1 full-parameter models.This study not only explores the application of large language models in the intellectual property domain,but also provides a reference for realizing the applicability of small parameter models in the judicial domain.

Named Entity Recognition for Chinese Based on Adaptive Attention and Boundary Enhancement

TANG Ruixue, WU Liqin, QIAN Qing

Computer Science. 2026, 53 (5): 309-318. doi:10.11896/jsjkx.250900076

Abstract ( 89 )

PDF(3617KB) ( 207 )

References | Related Articles | Metrics

NER(Named Entity Recognition) is a fundamental task in natural language processing,with extensive applications in information extraction,question answering systems,and knowledge graph construction.However,existing approaches still struggle with inadequate multi-scale feature utilization and inaccurate boundary identification when processing nested entities and ambiguous entity boundaries in Chinese text.To tackle these challenges,this paper proposes a Chinese NER model incorporating an AAM(Adaptive Attention Mechanism) and a BEM(Boundary Enhancement Module),specifically designed to handle the absence of explicit word delimiters and complex semantic structures in Chinese.The AAM dynamically integrates local and global contextual features to enhance the modeling of intricate Chinese semantic patterns,while the BEM employs depthwise convolution to strengthen boundary perception,effectively reducing recognition errors caused by nested entities and ambiguous spans.Experimental results demonstrate that the proposed model achieves F1 scores of 94.39% and 83.72% on the nested Chinese datasets ACE2005-Chinese and Cnerta,and 77.75%,84.88%,and 96.36% on the flat Chinese datasets Weibo,Ontonotes,and Resume,consistently surpassing existing mainstream Chinese NER methods and validating its effectiveness and generalization capability across diverse Chinese text scenarios.

Span-based Aspect Sentiment Triplet Extraction Based on Multi-view Graph Neural Networks

SHEN Ao, ZHOU Qingkai, XIA Tian, GAO Ruiling

Computer Science. 2026, 53 (5): 319-327. doi:10.11896/jsjkx.250200126

Abstract ( 80 )

PDF(2516KB) ( 196 )

References | Related Articles | Metrics

Aspect sentiment analysis triplet extraction(ASTE) is an emerging task in fine-grained sentiment analysis.Traditional span-level sentiment triple extraction methods have recently achieved remarkable results on the ASTE task,but these methods have not fully exploited the potential of syntactic and semantic dependencies.This study proposes a multi-view edge-enhanced encoder that makes full use of the syntactic and semantic relationships between words to accurately distinguish aspect words and viewpoint words and obtain rich deep information.Specifically,a dual-channel encoder with RoBERTa is first used to obtain basic semantic information,and at the same time,a bi-directional long short-term memory network channel and the edge-enhanced graph neural network are utilized to comprehensively capture semantic and syntactic information.Considering the insufficient sensitivity of traditional span-level models to span boundaries,it introduces a multi-layer graph convolutional network to capture the cross-relationships between spans to effectively identify span boundaries.In addition,this study also uses the mutual difference elimination strategy to eliminate conflicting triples.Through extensive experiments on multiple public benchmark data sets,this method outperforms other baseline models and verifies its effectiveness in the emotional triple extraction task.

Spatio-Temporal Trajectory Planning for Unmanned Vehicles in Complex Environments

ZHENG Yayu, RAO Pinyang, MU Jianbin, ZHU Wei

Computer Science. 2026, 53 (5): 328-336. doi:10.11896/jsjkx.250300043

Abstract ( 86 )

PDF(4697KB) ( 203 )

References | Related Articles | Metrics

To address the inefficiency of handling environmental constraints and kinematic constraints in trajectory planning for Ackermann-steering autonomous vehicles in complex environments,this paper proposes an efficient spatiotemporal joint trajectory planning algorithm with front-end and back-end coupling.In the front-end path planning,by integrating a topology map-guided sliding window pruning strategy with a dynamic heuristic function optimization method,the search efficiency is improved.For the back-end trajectory optimization,environmental constraints are first represented by dual safety corridors.Based on the differential flatness characteristics of the Ackermann model vehicle,kinematic constraints are expressed using the vehicle’s differentially flat variables and their higher-order derivatives as optimization variables.These variables are then analytically transformed into an unconstrained space via smooth mapping and diffeomorphic transformations.Finally,gradient optimization is performed in the unconstrained space to resolve constraint-handling inefficiencies.Experiments conducted in real-world parking environments demonstrate that the proposed method achieves 47% faster path search time compared to existing methods while maintaining comparable path quality.The trajectory optimization results show that the proposed algorithm can plan high-quality spatiotemporal trajectories in real time,verifying its effectiveness.

Study on Drug Target Affinity Prediction Based on Multi-view Comparison and Homology Information

TIAN Xin, ZHU Guosheng, XIONG Yuran, WU You

Computer Science. 2026, 53 (5): 337-345. doi:10.11896/jsjkx.250300168

Abstract ( 79 )

PDF(3158KB) ( 205 )

References | Related Articles | Metrics

Existing graph neural network(GNN)-based methods for drug-target affinity(DTA) prediction exhibit notable limitations in feature utilization and view alignment.Specifically,1) they fail to sufficiently model the feature representations of drugs and proteins within their respective views,leading to incomplete intra-view feature learning and restricted exploitation of feature information;2) they are unable to effectively align the intrinsic correlations across different views,thereby limiting cross-view information synergy.To address these challenges,a MVHGNN(Multi-View Hybrid Homogeneous Graph Neural Network ) is proposed.MVHGNN constructs a multi-view contrastive learning framework,employing an ESTGCN(Enhanced Subgraph Topology Graph Convolutional Network) and a GIN(Graph Isomorphism Network) as encoders in the drug molecular view and protein view,respectively,to capture the topological and hierarchical features of drugs and proteins.Furthermore,homology information is integrated to enhance intra-view feature representation and utilization.In the drug-target affinity view,a GCN(Graph Convolutional Network) is used to extract global topological information,enabling the construction of drug-protein interaction representations.A cross-view contrastive learning strategy is further adopted to maximize mutual information between drugs and proteins across different views,enhancing representation consistency and cross-view collaboration.Experimental results demonstrate that MVHGNN achieves superior performance on two benchmark datasets,notably reaching a mean squared error(MSE)of 0.166 and a modified determination coefficient(r²_m) of 0.794 on the Davis dataset,outperforming existing state-of-the-art methods.

Computational Complexity of Strictly d-regular (3,s)-SAT Problem

WANG Yongping, FU Zufeng

Computer Science. 2026, 53 (5): 346-353. doi:10.11896/jsjkx.250800047

Abstract ( 72 )

PDF(1567KB) ( 153 )

References | Related Articles | Metrics

he strictly d-regular (3,s)-SAT problem determines whether a given strictly d-regular (3,s)-CNF formula is satisfiable,where a CNF formula is satisfiable if there exists a truth assignment to its variables such that the truth value of the formula is true.A CNF formula is a 3-CNF formula if each clause contains three literals;a 3-CNF formula is a regular (3,s)-CNF formula if each variable appears s times in the formula;and a regular (3,s)-CNF formula is a strictly d-regular (3,s)-CNF formula if the absolute value of the difference between the positive and negative occurrences of each variable is d.Existing research shows that the strictly d-regular (3,s)-SAT problem has an aggregation phenomenon of hard-solving instances under some conditions.Hence,this paper studies the computational complexity of the strictly d-regular (3,s)-SAT problem.It shows that there exists both an unsatisfiable strictly (s-2)-regular (3,s)-CNF formula and an unsatisfiable strictly(s-4)-regular(3,s)-CNF formula when s≥4.Thus,the assumptions underlying the existing series of conclusions are confirmed.Hence,it determines the computational complexity of the strictly d-regular (3,s)-SAT problem as follows.A strictly d-regular (3,s)-CNF formula is satisfiable when s≤3;the strictly d-regular (3,s)-SAT problem is NP-complete when s≥4 and d<s;and a strictly d-regular (3,s)-CNF formula is satisfiable when s≥4 and d＝s.It is necessary to say that both the strictly 0-regular (3,4)-SAT problem and the strictly 2-regular (3,4)-SAT problem are NP-complete by the conclusion above.In other words,the strictly 0-regular (3,4)-SAT problem and the strictly 2-regular (3,4)-SAT problem inherit the NP-completeness of both the regular (3,4)-SAT problem and the 3-SAT problem.So,it should be necessary to study further the strictly d-regular (3,s)-SAT problem,especially,to study the strictly 0-regular (3,4)-SAT problem and the strictly 2-regular (3,4)-SAT problem.

Computer Architecture

Optimizing Distributed GMRES Algorithm with Floating-point Vector Compression

CHEN Shutong, GAO Jianhua, JI Weixing, LI Chunfeng

Computer Science. 2026, 53 (5): 354-366. doi:10.11896/jsjkx.250500033

Abstract ( 79 )

PDF(4844KB) ( 231 )

References | Related Articles | Metrics

The generalized minimal residuals(GMRES) method is one of the mainstream iterative algorithms for solving sparse linear systems,widely applied in circuit simulation,computational fluid dynamics,autonomous driving,and other fields.To efficiently solve large-scale sparse linear systems,researchers have developed distributed GMRES algorithms that use inter-node communication and distributed computing to achieve parallel solutions for large-scale problems.However,the sparse matrix-vector multiplication(SpMV) computation in distributed GMRES introduces significant communication demands for floating-point vectors,making the communication overhead of floating-point vectors a major performance bottleneck.Based on the restarted GMRES,this paper proposes a distributed restart GMRES(DR-GMRES) algorithm tailored for CPU-GPU heterogeneous systems,incorporating a collaborative computing model designed to align with the algorithm’s characteristics and heterogeneous architecture.Furthermore,for the communication requirements of floating-point vectors in SpMV and the numerical properties of GMRES’s inner-outer iterations,a two-layer hybrid compression strategy is proposed.This strategy combines lossy and lossless hybrid compression with inner-outer iteration hybrid compression to minimize the communication overhead of DR-GMRES while preserving acceptable solving accuracy.Additionally,a multi-buffer compressed communication architecture is designed to implement asynchronous data compression efficiently.Experimental results on 16 large-scale sparse linear systems demonstrate that,compared to the original distributed GMRES,the DR-GMRES algorithm based on hybrid data compression achieves an average speedup of ×3.37(with a maximum of ×8.02) with negligible impact on computational accuracy.Actual simulations experimentally verify that the DR-GMRES algorithm based on hybrid data compression achieves a ×3.09 acceleration in OSQP solving time over the baseline.

High-reliability Embedded Software Runtime Verification Method

LIU Jiale, HE Dongmei, LIU Hongbiao, PAN Guangyu, GUAN Yong, WANG Rui, LIU Bo

Computer Science. 2026, 53 (5): 367-375. doi:10.11896/jsjkx.251200184

Abstract ( 75 )

PDF(3473KB) ( 201 )

References | Related Articles | Metrics

High-reliability embedded software is widely deployed in safety-critical domains such as aerospace,industrial control,and automotive systems,often operating under harsh conditions including extreme temperature fluctuations,electromagnetic interference,and high-energy particle radiation.These conditions can easily trigger hardware-level anomalies like single-event upsets and data bit flips,leading to system functional failures or even major safety incidents.To enhance the runtime safety and reliability of high-reliability embedded software,this paper proposes a runtime verification method tailored for such software.Firstly,it performs static analysis on the target program to extract jump dependencies between instructions and constructs a priori instruction control flow information table.Secondly,it designs a lightweight monitoring module to capture real-time instruction flow data from the program’s bus,while verifying the correctness of the runtime instruction sequence through validation units combined with the a priori information table.Finally,it implements the verification framework on an FPGA platform and conducts experimental validation.The results demonstrate that the proposed method effectively ensures the correctness of high-reliability embedded software while meeting the resource constraints and verification latency requirements of embedded scenarios.

Improved Hippopotamus Algorithm for Energy Efficiency Optimization of HeterogeneousIntelligent Storage Computing

WANG Enliang, XIA Jun, SUN Zhixin

Computer Science. 2026, 53 (5): 376-387. doi:10.11896/jsjkx.250300140

Abstract ( 99 )

PDF(3844KB) ( 192 )

References | Related Articles | Metrics

The high energy consumption issue in data centers has become increasingly prominent amid digital transformation and the “dual carbon” strategy,stemming from energy efficiency bottlenecks caused by frequent data transmission in traditional distributed architectures.Existing optimization methods are confined to single dimensions(such as task scheduling or storage hierarchy optimization),struggling to collaboratively adapt to dynamic workloads and heterogeneous resource environments.This paper proposes an energy efficiency optimization framework for intelligent storage-computing systems based on the improved hippopo-tamus optimization algorithm(IHOA),which jointly optimizes task allocation and data placement through a multi-dimensional unified model encompassing computation,storage,and communication energy consumption.The framework innovatively introduces a storage-computing collaborative awareness mechanism that quantifies task-data correlations,combined with energy-efficiency-sensitive adaptive search strategies that dynamically adjust local and global search intensities to accommodate the energy efficiency characteristics of heterogeneous devices.Experimental results demonstrate that compared to mainstream optimization algorithms,IHOA significantly reduces total energy consumption in medium to large-scale systems,with efficiency improvements ranging from 8.1% to 25.6%,deriving its advantages from efficient suppression of remote data transmission energy consumption and dynamic adaptation to heterogeneous resources.Energy composition analysis further validates IHOA’s effectiveness in global collaborative optimization,achieving 17%~32% reduction in data transmission energy consumption by minimizing cross-node data migration.This research provides theoretical support and technical pathways for the green design of intelligent storage-computing systems,driving data centers toward high-efficiency,low-carbon development,while offering methodological references for energy efficiency optimization in emerging scenarios such as edge computing.

Novel Multi-task Federated Learning Based Approach for Detecting and Diagnosing Anomalies inCloud Microservices

CHEN Peng, HAO Junfeng, XIA Yunni, LI Xi

Computer Science. 2026, 53 (5): 388-403. doi:10.11896/jsjkx.250300131

Abstract ( 73 )

PDF(4650KB) ( 211 )

References | Related Articles | Metrics

Microservice architecture is widely used for application development in cloud environments,and its essence is to build applications through a series of functionally independent small autonomous services with high cohesion,high availability,low coupling,and good scalability.However,since microservice architecture is a distributed computing architecture with high dynamics and real-time system anomaly detection of distributed and independent microservices is a very challenging task,determining the category of detected anomalies is even more critical in practical applications.To solve the above problems,a multi-task federated learning-based system anomaly detection and diagnosis(MT-FL-SADD) method is proposed.Firstly,a multi-task federated lear-ning(MT-FL) distributed learning framework is proposed,which is used to construct an anomaly detection and diagnosis model for each microservice.Secondly,in order to identify the complex system anomaly patterns and features at the runtime of microservices,a feature extractor based on squeeze excitetion and external attention(SE-EA-EDN) is constructed to efficiently extract the features of real-time data from microservices monitoring at the runtime.Finally,a local-global feature-based parallel knowledge transfer(LGF-PKT) framework is designed to parallelize the weight update of local and global features.To validate the effectiveness of the proposed method,MT-FL-SADD improves the average Macro F1 by 33.9% and the average Micro F1 by 33.4% compared to other federated learning methods on the microservices benchmarking platforms Sock Shop and Train Ticket,and also improves the average F1 by 2.2% compared to other federated learning methods on SWaT,SMD and SKAB.

Information Security

Survey of Adversarial Sample Attacks for Vision Transformer

GUO Jingchen, YANG Kuiwu, DING Mengdi, WEI Jianghong

Computer Science. 2026, 53 (5): 404-418. doi:10.11896/jsjkx.250600065

Abstract ( 96 )

PDF(4019KB) ( 202 )

References | Related Articles | Metrics

Vision Transformer(ViT)is a novel architecture that breaks through the local receptive field limitation of traditional convolutional neural network and has made breakthrough progress in the field of computer vision with its global modeling capability.With the surging application of ViT in the security field,the structural differences between it and CNN have sharply reduced the effectiveness of traditional adversarial attacks,leading to the masking of the true vulnerability of ViT and the lag in the development of defense mechanisms.The model security risks triggered by adversarial attacks are driving research in this field to become a hot topic.This paper first systematically reviews the core progress of ViT adversarial attack methods,and analyzes the influence of ViT-specific structures such as image blocking,position coding,and attention mechanisms on adversarial sample attacks.Secondly,it classifies the adversarial attack methods for ViT,and divides the existing key attack methods into white-box attacks,migration-based black-box attacks,and decision-based black-box attacks.And it focuses on introducing the research progress of five types of black-box migration attacks,namely,optimization attacks on model structure,attacks based on input transformation,attacks based on integral gradient,attacks on downstream tasks,and attacks on model alignment.Then,the gradual evolution of different methods in terms of disturbance efficiency and cross-model migration is deeply explored.The core advantages and disadvantages of various attack methods are systematically summarized,revealing the evolution logic of attack techniques and model defects to provide references for the innovation of offensive and defensive technologies.Finally,the future research directions are analyzed and prospected.

Attack Capability Feature Learning and Aggregation Method Based on Semantic Co-occurrenceNetwork

LI Jingwen, ZHANG Ru, LIU Gongshen, ZHANG Tong

Computer Science. 2026, 53 (5): 419-425. doi:10.11896/jsjkx.250400070

Abstract ( 79 )

PDF(1796KB) ( 167 )

References | Related Articles | Metrics

The attack capability features of malicious samples can reflect the technical tactics of attackers,which is an important clue for advanced persistent threat(APT) attribution analysis.However,the continuous evolution of APT attack techniques tends to cause feature drift,making traditional fixed-feature-engineering methods difficult to adapt to new feature distributions and thereby weakening the model’s generalization ability.To address this issue,a robust attribution framework for attack capability features is proposed.Firstly,a semantic co-occurrence network of attack capability features is constructed,and an inductive graph neural network is introduced to jointly learn the semantics and co-occurrence relationship of features.Semantic clustering is then used to compress the feature space to generate a more stable aggregated feature representation.Secondly,a feature induction method is designed to realize the semantic generalization of unknown features,and a soft voting mechanism is used to integrate the predictions of multiple models,thereby improving the robustness and generalization capability of the attribution analysis.Experiments on a malicious sample dataset containing 91 APT groups show this method achieves macro-average and micro-average F1 scores of 71.46% and 81.15%,respectively,demonstrating its effectiveness.

Momentum Method with Monotonical Coordinate-wise Step-sizes for Adversarial Attacks

CHEN Jun, TAO Wei, BAO Lei, TAO Qing

Computer Science. 2026, 53 (5): 426-434. doi:10.11896/jsjkx.250600185

Abstract ( 81 )

PDF(3100KB) ( 188 )

References | Related Articles | Metrics

The generation of adversarial samples can be due to an optimization problem aimed at maximizing the objective functions of models.Currently,the strategies to solve the induced problems primarily rely on sign-gradient or sign-momentum methods.However,these approaches sacrifice critical gradient and momentum direction information,often leading to convergence issues and then resulting in instability of adversarial attacks.Inspired by the convergence analysis of AMSGrad,this paper proposes a momentum method with monotonical coordinate-wise step-size(MCS-MI) based on MI-FGSM,which enforces monotonically decreasing coordinate-wise step-sizes.For general convex cases,MCS-MI is proved to attain an optimal convergence rate of O(1/T),where T is the number of iterations.Furthermore,the strategy of enforcing monotonic coordinate-wise step-sizes is a general and efficient technique that can be integrated with existing momentum-based attack algorithms.Experimental comparisons with eight state-of-the-art adversarial attack methods on benchmark datasets demonstrate that the proposed approach not only exhibits superior stability but also significantly improves attack success rates,achieving maximum increases of 12.3% on CNN models and 5.9% on ViTs(Vision Transformers) respectively.

Technologies for Evaluating Defense Effectiveness of Endogenous Security Information Systems Based onAttack Graphs

CUI Tao, SHEN Junxia, CHEN Lin, ZHANG Yuntao, CHEN Monan

Computer Science. 2026, 53 (5): 435-445. doi:10.11896/jsjkx.250300130

Abstract ( 77 )

PDF(3825KB) ( 204 )

References | Related Articles | Metrics

With the increasing complexity and diversity of cybersecurity threats,traditional defense techniques are struggling to cope with evolving attack methods.Endogenous security technologies,especially those based on metamorphic defense,exhibit strong defense capabilities due to their dynamic adaptability,heterogeneity,and redundancy.This paper proposes an evaluation method for the defense effectiveness of endogenous security technologies based on attack graph modeling.By constructing network attack path models,the method quantifies the defense effects of endogenous security technologies in various attack scena-rios.Firstly,attack graph modeling is employed to describe network node vulnerabilities,attack paths,and their evolution,enabling the quantitative analysis of attacker behavior.Next,the impact of endogenous security technologies on attack paths is examined,with pre-implementation and post-implementation comparisons to assess defense effectiveness.The paper establishes a hierarchical security measurement framework,assessing the defense capabilities of inherent security technologies in terms of static defense at the node level,dynamic defense at the attack path level,and resilience recovery at the system level.Finally,simulation experiments demonstrate the effectiveness of the proposed evaluation method,providing a scientific basis for the quantitative evaluation of endogenous security technologies.

HEVC Information Hiding Algorithm Based on PU Partition Modes and Motion Vectors

YIN Hemin, ZHANG Yingnan, LI Jun, ZHANG Yanzhe

Computer Science. 2026, 53 (5): 446-452. doi:10.11896/jsjkx.251100040

Abstract ( 85 )

PDF(2998KB) ( 194 )

References | Related Articles | Metrics

With the increasing number of video users on Internet platforms,the high efficiency video coding(HEVC) standard is widely applied,steganography and steganalysis techniques based on the HEVC standard continue to develop.To enhance embedding security,this paper proposes an algorithm that utilizes prediction unit(PU) partition modes and motion vectors(MV) as carriers,dispersing embedding perturbation features while expanding the available embedding space.The PU partition mode domain defines binary mapping rules according to the quantitative structure inherent in the PU partition modes themselves,while the distortion function takes into account cumulative inter-frame errors,allowing the stego video to maintain the intrinsic distribution pattern of PU partition modes.In the motion vector domain,motion vector differences(MVD) that retain local optimality after preselected serve as carriers,with the distortion function considering both video quality and coding cost.Both domains employ syndrome trellis codes(STC) to adaptively find paths that minimise embedding disturbances.Experimental results demonstrate that the proposed algorithm maintains high video quality,exhibits strong resistance to steganalysis,and achieves an ideal embedding capacity.

[an error occurred while processing this directive]

2026,53 (5)

2026,53 (4)

2026,53 (3)

2026,53 (2)

2026,53 (1)

2025,52 (12)

2025,52 (11A)

2025,52 (11)

2025,52 (10)

2025,52 (9)

2025,52 (8)

2025,52 (7)

2025,52 (6A)

2025,52 (6)

2025,52 (5)

2025,52 (4)

More>>

Privacy Policy Compliance Detection Method for Mobile Application Based on Large LanguageModel (897)
WANG Limei, HAN Linrui, DU Zuwei, ZHENG Ri, SHI Jianzhong, LIU Yiqun
Computer Science. 2025, No.8:1-16

Abstract (897) PDF (3795KB) (23314)

Robust Hash Learning Method Based on Dual-teacher Self-supervised Distillation (4050)
MIAO Zhuang, WANG Ya-peng, LI Yang, WANG Jia-bao, ZHANG Rui, ZHAO Xin-xin
Computer Science. 2022, No.10:159-168

Abstract (4050) PDF (4472KB) (16053)

Data-free Model Evaluation Method Based on Feature Chirality (2627)
MIAO Zhuang, JI Shipeng, WU Bo, FU Ruizhi, CUI Haoran, LI Yang
Computer Science. 2024, No.7:337-344

Abstract (2627) PDF (3883KB) (15561)

Review of Time Series Prediction Methods (4779)
YANG Hai-min, PAN Zhi-song, BAI Wei
Computer Science. 2019, No.1:21-28

Abstract (4779) PDF (1294KB) (12917)

Polynomial Time Algorithm for Hamilton Circuit Problem (6697)
JIANG Xin-wen
Computer Science. 2020, No.7:8-20

Abstract (6697) PDF (1760KB) (12741)

Web Application Page Element Recognition and Visual Script Generation Based on Machine Vision (1248)
LI Zi-dong, YAO Yi-fei, WANG Wei-wei, ZHAO Rui-lian
Computer Science. 2022, No.11:65-75

Abstract (1248) PDF (2624KB) (12638)

Optimization Method of Streaming Storage Based on GCC Compiler (1732)
GAO Xiu-wu, HUANG Liang-ming, JIANG Jun
Computer Science. 2022, No.11:76-82

Abstract (1732) PDF (2713KB) (12319)

Research and Progress on Bug Report-oriented Bug Localization Techniques (1390)
NI Zhen, LI Bin, SUN Xiao-bing, LI Bi-xin, ZHU Cheng
Computer Science. 2022, No.11:8-23

Abstract (1390) PDF (2280KB) (11726)

Patch Validation Approach Combining Doc2Vec and BERT Embedding Technologies (922)
HUANG Ying, JIANG Shu-juan, JIANG Ting-ting
Computer Science. 2022, No.11:83-89

Abstract (922) PDF (2492KB) (11677)

Semantic Restoration and Automatic Transplant for ROP Exploit Script (986)
SHI Rui-heng, ZHU Yun-cong, ZHAO Yi-ru, ZHAO Lei
Computer Science. 2022, No.11:49-54

Abstract (986) PDF (2661KB) (11375)

Decision Tree Algorithm-based API Misuse Detection (1375)
LI Kang-le, REN Zhi-lei, ZHOU Zhi-de, JIANG He
Computer Science. 2022, No.11:30-38

Abstract (1375) PDF (3144KB) (11349)

Study on Effectiveness of Quality Objectives and Non-quality Objectives for Automated Software Refactoring (929)
GUO Ya-lin, LI Xiao-chen, REN Zhi-lei, JIANG He
Computer Science. 2022, No.11:55-64

Abstract (929) PDF (3409KB) (11322)

AutoUnit:Automatic Test Generation Based on Active Learning and Prediction Guidance (1521)
ZHANG Da-lin, ZHANG Zhe-wei, WANG Nan, LIU Ji-qiang
Computer Science. 2022, No.11:39-48

Abstract (1521) PDF (2609KB) (11263)

Study on Integration Test Order Generation Algorithm for SOA (1183)
ZHANG Bing-qing, FEI Qi, WANG Yi-chen, Yang Zhao
Computer Science. 2022, No.11:24-29

Abstract (1183) PDF (1866KB) (10646)

Studies on Community Question Answering-A Survey (731)
ZHANG Zhong-feng,LI Qiu-dan
Computer Science. 2010, No.11:19-23

Abstract (731) PDF (551KB) (10548)

Research Progress and Challenge of Programming by Examples (1127)
YAN Qian-yu, LI Yi, PENG Xin
Computer Science. 2022, No.11:1-7

Abstract (1127) PDF (1921KB) (9996)

Survey of Cloud-edge Collaboration (2735)
CHEN Yu-ping, LIU Bo, LIN Wei-wei, CHENG Hui-wen
Computer Science. 2021, No.3:259-268

Abstract (2735) PDF (1593KB) (9255)

Survey of Distributed Machine Learning Platforms and Algorithms (2443)
SHU Na,LIU Bo,LIN Wei-wei,LI Peng-fei
Computer Science. 2019, No.3:9-18

Abstract (2443) PDF (1744KB) (8976)

Survey of Fuzz Testing Technology (1984)
ZHANG Xiong and LI Zhou-jun
Computer Science. 2016, No.5:1-8

Abstract (1984) PDF (833KB) (8259)

Multisource Information Fusion:Key Issues,Research Progress and New Trends (644)
CHEN Ke-wen,ZHANG Zu-ping and LONG Jun
Computer Science. 2013, No.8:6-13

Abstract (644) PDF (746KB) (8147)

Featured ArticlesMore...