Computer Science

Research on Facial Emotion Expression Technologies for Humanoid Robots

KE Xianxin, LI Xuan, SONG Junqi

Computer Science. 2026, 53 (7): 1-8. doi:10.11896/jsjkx.250600166

Abstract

PDF(1861KB) ( 52 )

References | Related Articles | Metrics

Facial emotional expression in humanoid robots represents a cornerstone technology for enabling naturalistic human-robot interaction,exhibiting broad application prospects in sectors such as medical support,educational services,and social entertainment.This paper reviews the developmental trajectory of the field,delineating its evolution across three stages based on technological capability:foundational emotional expression,enhanced affective interaction,and personalized emotional resonance.It systematically examines the key breakthroughs within each stage concerning the design of biomimetic mechanical structures and the construction of multimodal affective interaction systems.The review places particular emphasis on the application of the facial action coding system,flexible biomimetic skin materials,deep learning,and multimodal large models(MLMs) in affective mode-ling.Furthermore,it investigates the critical role of multimodal spatio-temporal alignment mechanisms in the coordinated optimization of facial expressions,vocalizations,and physical actions.Finally,the paper outlines future research trajectories,including the fusion of high-precision actuation and sensing,the integration of multimodal large models,the enhancement of personalization and cultural adaptability,and strategies for mitigating the uncanny valley effect.This work provides a theoretical foundation to advance the intelligent and naturalistic development of facial emotional expression technology in humanoid robots.

Survey of Hyperbolic Geometry in Computer Vision

ZHU Yifei, LIU Tianpeng, SUN Tengzhong, LI Yanchen, CHEN Zhihong, FANG Pengfei

Computer Science. 2026, 53 (7): 9-23. doi:10.11896/jsjkx.250600134

Abstract

PDF(3724KB) ( 42 )

References | Related Articles | Metrics

Modeling data geometry within the learning community involves exploring the inherent structures among samples in a dataset,which is crucial for encoding the underlying data structures in the representation space.Hyperbolic geometry,characteri-zed as a Riemannian manifold with constant negative sectional curvature,offers a compelling alternative for embedding spaces across various learning scenarios,such as natural language processing and graph learning.This is largely due to its unique ability to encode hierarchical structures within data,such as those found in irregular graphs or tree-like datasets.Recent studies have also demonstrated the presence of hierarchical structures in visual datasets,and have explored the effective application of hyperbolic geometry within the computer vision(CV) sphere,ranging from classical image classification to advanced model adaptation lear-ning.This paper presents the most recent and comprehensive literature review on the application of hyperbolic spaces in CV.It introduces the fundamentals of hyperbolic geometry,followed by a thorough examination of algorithms that leverage the geometric priors of hyperbolic space in visual applications.These algorithms span unsupervised learning,supervised learning,and model adaptation learning.It concludes the manuscript by summarizing the findings and identifying potential future research directions.This article provides a clear overview of the practical advancements in hyperbolic geometry within the CV domain and aims to inspire further theoretical and practical developments in this field.

AETC:Image Classification Model via Attention-based Topological Features Fusion

ZHU Bin, LI Xiaobin

Computer Science. 2026, 53 (7): 24-33. doi:10.11896/jsjkx.250700003

Abstract

PDF(2866KB) ( 53 )

References | Related Articles | Metrics

A novel attention-enhanced topology and convolution model,which combines persistent homology and convolutional neural network through an attention-guided fusion framework,is proposed to address degradation in classification performance caused by insufficient topological feature extraction and weak intra-class structural consistency.Persistent homology captures key topological structures and encodes them into feature descriptors,while convolutional neural network extracts local visual features through convolution and pooling operations.An attention mechanism then merges both into a unified global representation to enhance feature completeness.The Wasserstein distance cross entropy loss function,derived by integrating the measurable Wasserstein distance with the cross-entropy loss,is used to constrain the topological structures of images.This effectively mitigates inter-class topological ambiguity,thereby enhancing the classification performance,robustness,and accuracy of the AETC model.Models with various topological vectorization methods are evaluated on three diverse datasets,the AETC model improves ACC by 2%~11%,AUC by 1%~7%,F1-score by 1%~11%,and mAP by 3%~17%.Within the classical convolutional neural network framework enhanced by persistence landscape vectorization,the optimal model achieves peak ACC of 95.49%,AUC of 99.44%,F1-score of 95.48%,and mAP of 98.42%.

FFiT:Faster Frame Interpolation Transformer Based on FasterViT

CHENG Zhirong, XU Yang

Computer Science. 2026, 53 (7): 34-44. doi:10.11896/jsjkx.250400046

Abstract

PDF(4212KB) ( 41 )

References | Related Articles | Metrics

In the realm of dynamic scene image acquisition,consumer-grade imaging devices commonly suffer from a compound degradation problem characterized by coexisting limited frame rates and motion blur.These issues primarily stem from inherent hardware architecture deficiencies and suboptimal exposure parameter optimization.Prevailing methodologies aimed at mitigating these degradations encounter significant technical bottlenecks in real-world blurry scenarios,including insufficient reconstruction accuracy,restricted model generalization capabilities,and low computational efficiency.To address these limitations,this paper introduces FFiT(Faster Frame Interpolation Transformer),a novel joint frame interpolation and motion decoupling optimization framework based on the FasterViT architecture.FFiT is designed to achieve efficient spatio-temporal joint modeling of dynamic blurry sequences.The framework integrates a CNN＋Transformer hybrid encoding architecture,incorporating several key mo-dules:1)An improved multi-scale residual transformer block(MRTB),which leverages spatio-temporal self-attention mechanisms to enhance reconstruction accuracy and explicitly model event correlations between blurry frames;2)A high-quality feature transmitter(HQFT) module,employing a cross-scale feature distillation mechanism to bolster semantic consistency during the blurry-to-sharp domain conversion,thereby addressing the challenge of model generalization;3)A lightweight dynamic upsampling and rendering module(DYRM) that utilizes differentiable dynamic convolution to decouple resolution reconstruction from computational complexity,thus tackling computational inefficiency and enabling flexible resolution recovery.Experimental evaluations on the Adobe240 and RBI datasets demonstrate that FFiT exhibits exceptional performance in addressing the intricate spatio-temporal joint modeling of dynamic blurry sequences.Notably,FFiT reduces model parameters by 70% to accommodate computational resource constraints while achieving a 2.73 dB improvement in PSNR(Peak Signal-to-Noise Ratio) compared to baseline models.This research offers a valuable reference for resolving such compound degradation problems by effectively balancing image quality enhancement with computational efficiency.

Image Anomaly Detection Based on Masked Convolutional Kernel

CHEN Yifan, DING Cong, CAO Min

Computer Science. 2026, 53 (7): 45-53. doi:10.11896/jsjkx.250900131

Abstract

PDF(3146KB) ( 18 )

References | Related Articles | Metrics

Image anomaly detection,which exploits patterns in normal images to detect abnormal images that do not conform to these patterns,plays a vital role in industrial production.Currently,one successful approach constructs a mask for normal images and uses a reconstruction network to predict the mask.The prediction error serves as the discriminant for image anomaly detection.However,this approach significantly increases model complexity when constructing the mask,compromising the reconstruction of normal image regions.To address this issue,this paper proposes a lightweight image anomaly detection method based on masked convolution kernels.This method applies a mask to the center of the convolution kernel and predicts the masked information through masked convolution and a channel-wise attention mechanism,effectively reducing the complexity of traditionalmas-king methods.A correction loss is introduced to constrain the reconstruction process of the masked predicted features,improving the reconstruction quality of normal regions.This method is validated on three industrial datasets MVTec,BTAD,and VISA,achieving AUROCs of 99.0%,93.7%,and 93.4%,respectively.This method can be combined with various existing reconstruction methods to effectively improve model performance without increasing model complexity.

Dual-view Separation and Reconstruction Method from Fused Random Compressed Measurement

HU Tao, CHEN Zan, FENG Yuanjing

Computer Science. 2026, 53 (7): 54-61. doi:10.11896/jsjkx.250400109

Abstract

PDF(3823KB) ( 12 )

References | Related Articles | Metrics

Compressed sensing(CS) technology has brought revolutionary advancements in image acquisition and reconstruction.However,research on multi-view CS is still in its early stages,and a unified optimization model for multi-view compressed reconstruction under single-sensor,single-measurement conditions has not yet been established.This paper proposes a CS framework tailored for dual-view scenarios,which effectively separates and reconstructs two distinct scene views from a single fused random measurement.The reconstruction task is decomposed into two sub-optimization problems and addressed using an iterative plug-and-play algorithm based on proximal gradient descent,incorporating image estimation and cross-view information interaction mechanisms.Dynamic message fusion is achieved through momentum feedback and residual adjustment.Experimental results demonstrate that,compared with other advanced single-view compressed sensing algorithms,the proposed method achieves higher reconstruction quality at low sampling rates,with a maximum PSNR improvement of 2.66 dB at a 10% compression ratio on the classic benchmark dataset Set11.

Low-light Image Enhancement Network Based on Multi-level Illumination Excitation and JointLoss Constraint

JIAO Hanbing, KANG Junhua, XIAO Teng, DENG Fei

Computer Science. 2026, 53 (7): 62-70. doi:10.11896/jsjkx.250400138

Abstract

PDF(4193KB) ( 11 )

References | Related Articles | Metrics

In computer vision applications such as autonomous driving and simultaneous localization and mapping(SLAM),low-light images frequently suffer from diminished contrast,noise interference,and detail loss,significantly impairing visual perception systems.Existing low-light enhancement methods exhibit limitations in noise suppression and color fidelity while demonstrating weak cross-scenario generalization capabilities.To address these challenges,this paper proposes a deep learning-based low-light enhancement approach using an improved Retinexformer architecture.The proposed method achieves effective enhancement through multi-stage feature excitation,global illumination adjustment,and multi-dimensional constrained optimization strategies.Firstly,it constructs MIFIB(Multi-level Illumination Feature Incentive Block) that enhances feature representation through normalization and an advanced channel attention mechanism,strengthening illumination modeling.Secondly,it designs a globally-aware IAB(Illumination Adjustment Block) to optimize illumination distribution in enhanced images.Finally,it introduces a multi-dimensional joint loss optimization strategy incorporating structural similarity constraints,semantic feature constraints,and color intensity consistency constraints to comprehensively guide model learning.Experimental results demonstrate that the proposed method achieves superior performance over state-of-the-art methods on the LOL benchmark in metrics including PSNR and SSIM.Generalization tests on SYNTHIA and Terrasentia datasets further validate the proposed method's advantages in noise suppression,color preservation,and detail retention.Moreover,quantitative evaluation in stereo matching-based 3D reconstruction shows that the proposed method reduces endpoint error(EPE) by 0.5 pixels－a 22.1% improvement compared to non-enhanced methods－confirming the advantages of the proposed method in maintaining scene geometric consistency and improving the robustness of depth perception.

Frequency-augmented and Multi-level Feature Fusion for Image-Text Sentiment Analyzer

ZHU Yuchao, ZHANG Shunxiang, WEN Boyu, SUN Liang, XU Yang

Computer Science. 2026, 53 (7): 71-79. doi:10.11896/jsjkx.250900117

Abstract

PDF(3742KB) ( 15 )

References | Related Articles | Metrics

Image-text sentiment analysis accurately mines users' emotional tendencies by utilizing the complementary information of text and images.Existing research mostly focuses on deep semantic feature interactions,and underutilizes shallow detailed features of the image-text modality,resulting in a decrease in the accuracy of sentiment polarity recognition.To address this pro-blem,this paper proposes a image-text sentiment analysis model based on frequency domain enhancement and multi-level feature fusion.Firstly,multi-level featuresfrom the image-text modality are extracted through the feature extraction module,covering different levels of features from shallow details to deep semantics.Secondly,in the frequency domain enhancement module,the high-frequency details of the shallow features and the low-frequency semantics of the deep features are enhanced through the synergistic optimization of the Fourier Transform and the coordinate attention mechanism,so as to improve the ability of different levels of feature characte-rization.Then,in the bimodal synergistic interaction module,through bidirectional attention and dynamic calibration,it effectively realizes the bidirectional cross-modal interaction of multi-level image and text features.Finally,the image and text features are fused layer by layer in the progressive fusion module,which realizes the effective utilization of the shallow detailed features to the deep semantic features.Experiments conducted on the dataset MVSA and HFM verify the effectiveness of the proposed model.

Research on Speed Driven Scene Scaling in Virtual Reality to Improve Walking and Searching Efficiency

JI Hongru, REN Yangfu, XU Senzhe, ZHANG Songhai

Computer Science. 2026, 53 (7): 80-90. doi:10.11896/jsjkx.250800026

Abstract

PDF(2913KB) ( 10 )

References | Related Articles | Metrics

Virtual reality(VR)provides users with an immersive 3D digital environment,but it is still a technical challenge to realize natural walking without being limited by physical space.Previous VR mobility methods,such as transmission,walking in place,and redirected walking(RDW),have made some progress in overcoming this challenge.However,these methods also face some limitations,such as the sense of unnatural,additional hardware requirements or the risk of motion sickness.This paper proposes an innovative VR movement method called “three-dimensional space contraction”,which is inspired by the Lorentz contraction phenomenon in special relativity.Similar to Lorenz contraction,the three-dimensional space contraction shrinks the three-dimensional virtual space according to the size of the user's moving speed,that is,when the user's speed is high,the virtual space shrinks more,at this time,the person will be relatively high,thus the vision will be wider,while when the user's moving speed is low,the three-dimensional virtual space shows almost no contraction.Using the method of three-dimensional space contraction,users can not only effectively span a long virtual distance through a shorter physical walk,but also have a broader vision.The difference from the translation gain is that space contraction effect is observable by users and consistent with their intentions,so there is no inconsistency between the user's ontology and visual perception.VR users will become giants according to the change of speed during walking,so that the vision will be wider and the search efficiency will be improved.When the user's physical movement speed is relatively slow,the virtual space approaches its original size,thereby reducing perceptual conflicts.3D space contraction is a general moving method,which has no special requirements for VR scenes.Experimental results of field user research in various virtual scenes show that 3D space contraction has a significant effect in improving walking efficiency and search efficiency.In addition,experiments also show that 3D space contraction has the potential to integrate with existing mobile technologies(such as RDW).

Identification of Authentic and Forged Paper-based Fingerprints Based on LG-GFNet Feature Fusion Network

NING Shiqiang, ZHOU Lianzhen, ZHANG Lifeng

Computer Science. 2026, 53 (7): 91-100. doi:10.11896/jsjkx.260300086

Abstract

PDF(4234KB) ( 17 )

References | Related Articles | Metrics

To address the challenge of accurately identifying high-fidelity silicone forged fingerprints in traditional morphological inspection,a local-global gated feature-level fusion network is proposed under the transfer learning framework for the automatic qualitative inspection of authentic and forged paper-based fingerprints.Aiming at the limitation that traditional recognition me-thods struggle to balance micro-texture and macro-smearing,the network takes a modern lightweight convolutional model as the backbone to extract stable local ridge features,while integrating a Patch convolution branch to capture cross-scale global morphological differences and consistency variations in ink diffusion.Adaptive fusion of local and global features is achieved through a gating mechanism.Experiments are conducted on 16 000 authentic and forged fingerprint samples collected from 20 volunteers using four types of media:red,blue,and black inkpad oils,as well as red inkpaste.The results demonstrate that the network maintains stable and high-precision recognition performance under multi-media and cross-individual conditions,with core metrics including accuracy,F1-score,and AUC overall outperforming traditional pattern recognition methods.Grad-CAM visualization results confirm that the model mainly focuses on regions such as ink diffusion boundaries,ridge fractures,and gray-scale abnormal bands,and its decision-making logic is highly consistent with the inspection experience of judicial examiners.This effectively improves the ability to distinguish subtle forged traces under complex media,providing a high-precision and interpretable technical approach for the intelligent detection of forged fingerprints in judicial identification.

Retrieval-Augmented Generation:Survey of Methods and Applications

WANG Xinlin, LI Yan, MA Chaofan, LI Shuo

Computer Science. 2026, 53 (7): 101-117. doi:10.11896/jsjkx.251000089

Abstract

PDF(2517KB) ( 12 )

References | Related Articles | Metrics

Large Language Model(LLM),endowed with human-like capabilities in text generation and natural language understanding,are profoundly reshaping the landscape of artificial intelligence.However,due to the closed nature and lagged updates of their training corpora,LLM often generate responses that deviate from factual accuracy when confronted with dynamically updated information or long-tail knowledge in specialized domains.To address this limitation,Retrieval-Augmented Generation(RAG) extends the knowledge scope of LLM by integrating external knowledge bases,thereby assisting them in producing more accurate and higher-quality answers.Distinct from prior classifications based on processing stages,this paper innovatively categorizes existing RAG approaches around core challenges to be addressed,dividing them into four major categories:chunk optimization,retrieval enhancement,context compression,and knowledge graph integration.This paper systematically reviews representative works and technical mechanisms within each category.Subsequently,it summarizes typical applications of RAG across various specialized domains,and finally discusses future research directions,providing researchers with a clear technical roadmap and practical guidance for method selection.

From Boolean Retrieval to Foundational Models:Legal Perspective on Development of Case Retrieval Technology

SHI Yiran, ZHANG Linghan, LIU Yiqun

Computer Science. 2026, 53 (7): 118-124. doi:10.11896/jsjkx.250600009

Abstract

PDF(1614KB) ( 11 )

References | Related Articles | Metrics

Technological innovation,exemplified by artificial intelligence,is reshaping judicial practice and driving the development of “smart courts”.In this context,similar case retrieval technology,as a key application of AI,is evolving from traditional keyword-based searches and shallow text matching to intelligent retrieval based on deep semantic understanding,contextual awareness,and logical reasoning.AI-powered retrieval can partially simulate judges' reasoning processes,improving efficiency,promoting consistency in legal application,and enhancing judicial fairness.However,in-depth interviews with frontline judges reveal a significant gap between the technology used in practice and the advancements seen in research.This gap manifests in three main respects:1)Etrieval methods remain predominantly keyword-based,with semantic-level retrieval yet to be widely adopted,resulting in limited efficiency.2)Retrieval quality remains inadequate,as existing databases struggle to balance case volume with case authority,and a unified and explicit standard for defining “similar cases” is still lacking.3)Retrieval rules are incomplete,with insufficient differentiation based on jurisdiction,trial level,and the legal effect of judgments.Drawing on the application of mainstream retrieval platforms,this paper adopts a mixed method combining cross-platform comparison with longitudinal technical analysis.From both legal theory and judicial practice perspectives,it examines the current development and existing bottlenecks of similar case retrieval technology and proposes directions for future improvement.

Event Causal Data Augmentation Method Based on Large Language Model

CHEN Zhixiang, XIE Zhipeng

Computer Science. 2026, 53 (7): 125-131. doi:10.11896/jsjkx.250600098

Abstract

PDF(2097KB) ( 22 )

References | Related Articles | Metrics

Event causality identification(ECI) is an important NLP task that aims to identify causal relationships between two events.However,due to the scarcity of causal data in public datasets,downstream ECI models have encountered bottlenecks.To alleviate the data scarcity problem,this paper proposes a large language model-based event causality data augmentation method(LLM-ECIAug).This method constructs a data generation strategy from two levels:causal event pairs and causal patterns.It utilizes large language models to generate diverse candidate augmentation data and combines an event causality filter fine-tuned on the original ECI datasets.In view of the differences between the distribution of candidate augmented data and the original data,a filtering mechanism based on KL divergence is introduced to rank and filter the generated data,retaining high-quality data that are most consistent with the original data distribution.Finally,the filtered augmented data is combined with the original data to train the downstream ECI model.Experimental results show that this method achieves better F1 scores than other data augmentation baseline methods on the Causal-TimeBank and EventStoryLine datasets,confirming its effectiveness and superiority.

Knowledge-enhanced Text Embedding Optimization Based on Key Information Extraction

CHE Yunli, TANG Jintao, WANG Ting, ZHANG Jian

Computer Science. 2026, 53 (7): 132-138. doi:10.11896/jsjkx.250600021

Abstract

PDF(2631KB) ( 15 )

References | Related Articles | Metrics

To address the challenges of semantic dilution and fine-grained information loss in long-text retrieval,this study proposes a text embedding optimization framework based on explicit knowledge extraction.Traditional methods that rely on single-vector representations of entire texts struggle to capture the multi-topic and hierarchical local semantic associations present in lengthy documents.It designs two knowledge-aware embedding strategies:1) Knowledge-aware separate embedding(KASE),which employs a large language model to extract key knowledge points from text segments and vectorizes them independently to preserve fine-grained semantics;2) Knowledge-aware concatenated embedding(KACE),which concatenates the extracted know-ledge points into a single passage and encodes it holistically to explore aggregation effects.Experimental results on Chinese datasets(CMRC,DRCD) and English datasets(SQuAD,NewsQA) demonstrate that the proposed knowledge-enhanced methods significantly outperform both traditional baselines and recent approaches such as QA-pair-augmented retrieval(QAEA-DR) and Hypothetical document embeddings(HyDE).KASE achieves the most substantial improvements,especially in long-text scenarios.Ablation studies reveal that independent knowledge-point representations are crucial for mitigating information loss,while hybrid strategies that combine original text embeddings with knowledge-aware vectors further optimize retrieval effectiveness.

Gate-controlled Agent Attention Mechanism-based Soft Prompt Transfer Method

ZHANG Yan, ZHOU Jian, HAN Lei, CHENG Chunling

Computer Science. 2026, 53 (7): 139-145. doi:10.11896/jsjkx.250600038

Abstract

PDF(1928KB) ( 12 )

References | Related Articles | Metrics

Soft prompt transfer,as a cross-task learning approach,aims to guide the target task model to learn more generalizable prompt representations by leveraging knowledge from source prompts.However,existing methods often ignore the impact of task-specific information embedded in target samples,which may result in biased target prompt training.To address this issue,a gate-controlled agent attention mechanism-based soft prompt transfer method(GPAPT) is proposed.Firstly,to separate and enhance both global and task-specific information in target samples,a dual-path feature extraction strategy is introduced.It employs two lightweight extraction paths to derive agent tokens that encode task-level semantics and local features that capture instance-specific characteristics.Secondly,to assess the transferability of source prompts,a gate-controlled agent attention mechanism is presented.It computes attention distributions between prompt features and both agent tokens and local features to model task similarity.To preserve task-specific information,a gating unit filters local features and integrates the two types of attention.Finally,a perturbation mechanism for attention distribution is introduced.It perturbs the attention weights based on the prediction loss of source prompts on the target task,thereby reduces the influence of source prompts that are similar in the representation space but perform poorly in prediction,and thus improving transfer robustness.Extensive experiments on the GLUE benchmark demonstrate that GPAPT achieves the best average performance compared to nine strong baseline methods,validating the effectiveness of the proposed approach.

Obstacle Avoidance Motion Planning of Robotic Arm Based on Improved RRT Algorithm

YANG Ming, ZHU Xuejun, LAI Huige, YU Checao, XIONG Leilei, PENG Da, MAO Kun

Computer Science. 2026, 53 (7): 146-155. doi:10.11896/jsjkx.250400044

Abstract

PDF(5692KB) ( 13 )

References | Related Articles | Metrics

In order to improve the success rate and efficiency of obstacle avoidance motion planning of cooperative robotic arms in complex environments such as cooperative operation and dynamic obstacles,a path planning method based on improved rapidly-exploring random tree(RRT) algorithm is proposed.On the basis of bidirectional RRT(Bi-RRT),the target bias strategy is first introduced.By reducing the randomness of the search process,the sampling efficiency can be improved.Secondly,in the process of random tree expansion,when the target point is not selected as a new node,the target point guidance strategy is adopted to promote the growth of the random tree in the direction of the target point,so as to effectively shorten the search time.Finally,the improved artificial potential field(APF) method is integrated into the improve Bi-RRT.By redesigning the potential field function and introducing the distance influence factor,it is endowed with local planning ability,reduces path redundancy,and further improves planning efficiency.Simulation results in two-dimensional and three-dimensional environments show that the proposed algorithm can generate shorter paths,and the random tree growth is more goal-oriented,and the planning time is significantly reduced.The algorithm is applied to the visual simulation of the robotic arm model,and the results show that it can effectively guide the robotic arm to avoid obstacles and accurately reach the target point.The effectiveness and practicability of the proposed methodare further confirmed by practical verification on the collaborative robotic arm.

Document-level Event Argument Extraction Model Based on Hierarchical Dependency Aggregation and Event Enhancement

DING Zhijun, AI Fangju, LIU Aihuan

Computer Science. 2026, 53 (7): 156-167. doi:10.11896/jsjkx.250500006

Abstract

PDF(2393KB) ( 15 )

References | Related Articles | Metrics

In the task of document-level event argument extraction,existing prompt-based learning models face challenges in comprehensively capturing the holistic semantic understanding of documents and neglect the interaction between documents and prompt templates.To address these limitations,a novel document-level event argument extraction model is proposed,incorporating hierarchical dependency aggregation and an event enhancement mechanism.Initially,the vector representations of the input sequence are obtained through BART.Subsequently,a hierarchical dependency aggregation architecture is introduced,consisting of a semantic hierarchical dependency-aware mechanism and a structural hierarchical dependency-aware mechanism,designed to capture semantic correlations and structural information among entity mentions respectively.These features are dynamically fused via a self-attention mechanism.The integrated relation matrix is then leveraged to guide the aggregation of contextual information from relational dependency graphs onto corresponding entity mentions,thereby updating mention representations.These enhanced representations are subsequently incorporated into the original input representations to strengthen the model's capability in perceiving entity relations and facilitating document-prompt template interactions.Finally,an event enhancement mechanism is implemented to prioritize contextually relevant information for the current extraction task while mitigating information redundancy and interference,ultimately improving argument extraction accuracy.Experimental results conducted on RAMS and WikiEvent two public dataset demonstrate that the proposed model achieves superior F1 scores compared to baseline models,with significant improvements in comprehensive performance metrics.

Application of Multi-strategy Improved Hippopotamus Optimization Algorithm in Path Planning

ZHANG Jiawei, MA Zhanyou, DING Beibei

Computer Science. 2026, 53 (7): 168-177. doi:10.11896/jsjkx.250500040

Abstract

PDF(4821KB) ( 12 )

References | Related Articles | Metrics

In view of the defects of the hippopotamus optimization algorithm,such as being prone to falling into local optima and having low search accuracy when solving path-planning problems,an improved hippopotamus optimization(IHO) algorithm with multiple strategies is proposed.Firstly,the Latin hypercube sampling method is introduced to initialize the hippopotamus population,which expands the initial exploration scope and enables the algorithm to be more uniformly distributed in the search space.Secondly,a dynamic adaptive convergence factor is introduced to improve the position update method of male hippopotamuses,enhancing the global search ability and reducing the probability of the algorithm falling into local optima.Then,a variable spiral search strategy is introduced in the hippopotamus defense phase to balance the exploitation and exploration abilities of the algorithm and improve the search efficiency.Simulation results of 12 benchmark test functions show that the IHO algorithm has better optimization ability and faster convergence speed compared with the hippopotamus optimization algorithm,grey wolf algorithm,sand cat swarm algorithm,and genetic algorithm.Finally,the IHO algorithm is applied to the path planning of mobile robots.Experimental results show that in 15×15,20×20,and 30×30 grid maps,the IHO algorithm shortens the path by 7.7%,1.8%,and 4.8% respectively compared with the hippopotamus optimization algorithm.It shows significant performance advantages.

Competitive Artificial Bee Colony Algorithm for Fuzzy Distributed Scheduling

ZHENG Youlian

Computer Science. 2026, 53 (7): 178-185. doi:10.11896/jsjkx.250400001

Abstract

PDF(1879KB) ( 9 )

References | Related Articles | Metrics

Distributed scheduling is important part of distributed manufacturing and the focus of the scheduling research.Uncertainty is the basic feature of distributed maufacturing.To solve fuzzy distributed two-stage hybrid flow shop scheduling problem with factory eligibility and additional resources,a competition artificial bee colony(CABC) algorithm is proposed to minimize makespan and total agreement index.To obtain high quality solutions,the whole population is divided into two bee swarms and competition process between them is constructed.In the competition process,the winning swarm is used as employed bee swarm and another one acts as onlooker bee swarm.The competition process is composed of employed bee phase and extra search,and adaptive onlooker bee phase.Five search strategies are given and the diversified search is implemented.An adaptive population reconstruction is applied to avoid excessive competition between swarms.Experiments are conducted and the computational results reveal that new strategies of CABC are effective and CABC has the promising search advantages on solving the considered pro-blem.

Multi-clause Dynamic Deduction Algorithm Based on Deduction Body

ZHOU Jie, ZOU Weigang, CAO Feng, YI Jianbing

Computer Science. 2026, 53 (7): 186-194. doi:10.11896/jsjkx.250400099

Abstract

PDF(1742KB) ( 8 )

References | Related Articles | Metrics

Given the synergized reasoning effect of path information on subsequent deduction,exploring effective representation and application methods of deduction paths holds significant academic value.In response to the multi-clause deduction process,this paper proposes the definition of deduction body,which systematically describes the deduction path information at each step of multi-clause deduction.According to the variation of remaining literals during deduction,a classification method,a representation method and a deduction method for deduction body are introduced.This method can generate and store different types of deduction bodies according to the deduction process,and utilize them to further eliminate the remaining literals in subsequent deduction steps.Based on this method,a multi-clause dynamic deduction algorithm grounded in deduction body is presented,in which unit clauses and candidate clauses are selected for deduction through weight-based ranking.The participating clauses are used to construct standard contradiction,whose effectiveness is then evaluated.If the deduction is valid,a deduction body is constructed;if the deduction is invalid,the clauses involved in the deduction are readjusted through a backtracking mechanism.After generating a standard contradiction separation clause,the deduction body is used to further reduce or enrich the deductive search path.This algorithm is applied in the internationally leading theorem prover Eprover3.2 and tested on benchmark FOF division problems from the 2023 and 2024 international automated theorem provers competitions.Within the testing time of 300 seconds,Eprover3.2with the proposed algorithm both solves 14 more problems than the original Eprover3.2 respectively.Furthermore,the problems from TPTP(Thousands of Problems for Theorem Provers) problem library with a rating of 1 are also used as a test object,Eprover3.2 with proposed algorithm solves 7 problems that can not be solved by all other provers.Experimental results demonstrate that the multi-clause dynamic deduction algorithm based on deduction body can be effectively applied to first-order logic automated theorem proving.

Feature Selection Based on Hierarchical Quantum Bat Algorithm

ZHANG Xingwang, HE Xiaoli, CHEN Si, SHE Yanhong

Computer Science. 2026, 53 (7): 195-204. doi:10.11896/jsjkx.250500032

Abstract

PDF(3901KB) ( 8 )

References | Related Articles | Metrics

Feature selection(FS),as a core step in pattern recognition,aims to optimize classification performance and computational efficiency through dimensionality reduction.To address the challenges faced by existing swarm intelligence algorithms,such as the BA(Bat Algorithm),in high-dimensional data,including low search efficiency and susceptibility to local optima,a IRHQBA(Hierarchical Quantum Bat Algorithm Based on Information Gain Ratio and Random Forest) is designed.Firstly,a hybrid filter-based pre-selection mechanism is constructed,integrating Pearson correlation,IGR(Information Gain Ratio),and RF(Random Forest) for triple evaluation,to efficiently eliminate redundant features.Secondly,in the BA initialization stage,hierarchical sub-feature grouping is conducted based on feature importance ranking,and quantum computing is introduced to optimize the search space and mutation strategy to enhance diversity.Finally,the classification performance of feature subsets is improved through hierarchical collaborative optimization and dynamic mutation mechanisms.

Traffic Flow Forecasting Based on Dynamic Graph Convolution and Hypergraph Learning

ZHAO Xingbo, LIAN Defu

Computer Science. 2026, 53 (7): 205-212. doi:10.11896/jsjkx.250600133

Abstract

PDF(2363KB) ( 11 )

References | Related Articles | Metrics

The traffic flow prediction problem aims to accurately forecast future traffic conditions based on historical observation data.In order to better model the complex spatiotemporal relationships in the traffic network,it is essential to leverage various correlation patterns embedded in the data,including pairwise correlations and higher-order correlations.However,mining these correlation patterns from traffic spatiotemporal data poses significant challenges.Unlike many other domains where prior know-ledge can be used to construct effective correlation structures,traffic data often lacks clear predefined relationships.The dynamic nature of traffic networks further complicates the task,as correlation patterns can change over time due to factors such as weather conditions,accidents,or special events.Therefore,developing methods that can automatically discover and adapt to these changing patterns is crucial for improving the accuracy of traffic flow predictions.In this paper,a novel approach is proposed to address these challenges by utilizing both graph and hypergraph structures.In the pairwise correlation modeling,node embeddings are utilized to infer the graph structure.To better capture the dynamic changes in correlations within the data,a dynamic graph structure learning method is designed to capture correlation patterns from the dynamically changing data.In the higher-order correlation modeling,matrix optimization is employed to directly learn the hypergraph structure,and a contrastive learning loss function is introduced to effectively capture global dependencies.In the temporal dimension,a temporal Transformer architecture is adoptedto model the time-dependent aspects of traffic flow.The self-attention mechanism,a key component of the Transformer,is utilized to achieve multi-scale modeling of temporal dependencies.This allows the model to consider both short-term and long-term patterns in the data,which are essential for accurate traffic flow prediction.Experiments on multiple commonly used traffic flow datasets show that the proposed model can effectively model the complex spatiotemporal relationships in the traffic network,and its prediction accuracy is significantly improved compared to various baseline models.

Causal Subgraph Learning for Cascade Popularity Prediction

LI Kaiju, YIN Chenyang, CHENG Zhangtao, LIU Xueting, ZHOU Fan

Computer Science. 2026, 53 (7): 213-221. doi:10.11896/jsjkx.250700055

Abstract

PDF(3015KB) ( 10 )

References | Related Articles | Metrics

Information cascade popularity prediction is critical for understanding the dynamics of information dissemination and mitigating the spread of misinformation.Although existing deep learning methods have achieved improvements in predictive accuracy,they still exhibit limitations in disentangling confounding factors and uncovering the deep causal relationships between cascade structures and popularity,which undermines both the reliability and interpretability of predictions.To address these challenges,this paper proposes a novel cascade popularity prediction model based on causal subgraph learning,named causal-aware cascade model(CauCas).CauCas introduces graph data augmentation to impose interventions on the original cascade graphs and encodes multi-level cascade representations for both the original and augmented graphs.Through specialized layer-wise feature selection and weighted fusion strategies,the model derives graph-level representations for each graph,and leverages adaptive instance normalization to learn robust features that are less sensitive to interventions and more likely to reflect causal relationships.Finally,the fused feature representations are fed into a multilayer perceptron to perform popularity prediction.Experimental results on Twitter,Weibo,and APS three public datasets demonstrate that CauCas achieves superior performance,consistently outperforming state-of-the-art methods across diverse datasets and prediction windows.

Meteorological-Personal Health Data Fusion Methods for Chronic Disease Prediction

HE Zhiguang, TAN Benchao, YU Hong, WANG Guoyin, LU Jiawei

Computer Science. 2026, 53 (7): 222-229. doi:10.11896/jsjkx.251100060

Abstract

PDF(2846KB) ( 8 )

References | Related Articles | Metrics

The environment exerts a prolonged influence on human health and the progression of diseases.Therefore,investigating the impact of climatic conditions on individual health and improving the accuracy of disease prediction models is of great significance for disease prevention and control.Currently,research integrating meteorological data with disease studies predominantly focuses on qualitative analysis or macro-level predictions at the population level,while quantitative modeling addressing indivi-dual differential responses remains relatively scarce.To address this issue,this paper proposes a chronic disease prediction method based on interaction models that integrates meteorological data with individual health data.The method utilizes data from the CLHLS-HF(Chinese Longitudinal Healthy Longevity and Happy Family Study) and theCHARLS(China Health and Retirement Longitudinal Study),employing Moran's I statistic to identify geographically distributed diseases with spatial correlation.Interaction models are used to integrate individual health data and meteorological data,constructing interaction features to enhance the performance of disease prediction models.Four machine learning models—logistic regression,Naive Bayes,XGBoost,and multilayer perceptron(MLP)—are applied to both datasets for disease prediction to evaluate the impact of incorporating meteorological features.Experimental results show that in the CLHLS-HF dataset,the Naive Bayes model improves specificity by 10.2% for dyslipidemia prediction,while the XGBoost model improves accuracy and sensitivity by 5.6%.In the CHARLS dataset,the multilayer perceptron model improves the AUC for heart disease prediction by 6.6%,and the XGBoost model improves sensitivity by 6.2%.These findings demonstrate that incorporating meteorological data and interaction features consistently enhances the performance of disease prediction models.

CBT-AD:CNN-BiLSTM-Transformer Hybrid Model for Time Series Anomaly Detection

XU Jian, CHEN Shijie, FENG Jiancong, YANG Geng

Computer Science. 2026, 53 (7): 230-241. doi:10.11896/jsjkx.250600078

Abstract

PDF(4334KB) ( 11 )

References | Related Articles | Metrics

Existing deep learning-based time series anomaly detection methods exhibit limitations in both the collaborative mode-ling of local and global features,as well as in computational efficiency for long sequences.To address these issues,this paper proposes a hybrid-architecture anomaly detection model.Firstly,for multi-scale feature extraction,depthwise separable convolution is employed to capture local detail features,combined with adaptive pooling to enhance the response capability to burst anomalies.Secondly,a bidirectional temporal modeling mechanism is designed,which integrates bidirectional context features through Bi-LSTM and utilizes gated Dropout to mitigate the risk of overfitting during long-sequence training.Furthermore,a hierarchical sparse global attention module is designed,leveraging local window attention to capture time-frequency features,while employing multi-head mechanisms and residual connections to optimize gradient propagation stability.Finally,a dynamic feature fusion method is proposed and integrated with a chunk-based processing framework to achieve collaborative optimization of detection accuracy and computational efficiency.Experimental results on four public time-series datasets demonstrate that the proposed model achieves significant improvements across various performance metrics compared to existing methods,while also exhibiting strong robustness and generalization capability.

Research on Modeling and Scheduling Methods for Intra-city Delivery Based on Heterogeneous Graph Neural Networks

WU Kai, SUN Zhe, ZHANG Xu, CAO Yadong, SUN Zhixin

Computer Science. 2026, 53 (7): 242-250. doi:10.11896/jsjkx.250400121

Abstract

PDF(2281KB) ( 7 )

References | Related Articles | Metrics

To address the challenges of heterogeneous feature modeling and insufficient interaction fusion in intra-city delivery scenarios,this paper proposes a Prompt-guided bidirectional heterogeneous graph Transformer model(P-BiHGT).The model introduces virtual Prompt nodes as global semantic anchors and explicitly connects them with both vehicle and order nodes,thereby enhancing the fusion and propagation of global semantic information across the graph.Furthermore,to tackle the weak semantic association and unidirectional interaction modeling between vehicle and order nodes,a role-aware bidirectional attention mechanism is designed to model interaction paths in both directions:from vehicles to orders and vice versa.After completing bidirectionalinteraction modeling,a multilayer perceptron(MLP) is employed to make high-level matching decisions on the fused feature pairs,improving the overall matching accuracy.Experimental results on a simulated intra-city delivery dataset show that the proposed model achieves a validation accuracy of 93.6%,significantly outperforming traditional models,thus demonstrating the effectiveness and adaptability of P-BiHGT in heterogeneous matching tasks.

iDSRformer:Node Load Prediction Model for High-performance Computing Cluster

XIAO Yanxue, DENG Li, REN Zhengwei, WU Mengxin

Computer Science. 2026, 53 (7): 251-261. doi:10.11896/jsjkx.250600026

Abstract

PDF(3075KB) ( 10 )

References | Related Articles | Metrics

High-performance computing(HPC) clusters play a crucial role in large-scale computing tasks,however,as the scale of clusters continues to expand,the management of cluster energy efficiency has become increasingly important.Prediction technology can provide decision support for energy efficiency management through accurate estimation of cluster resource usage,thereby achieving dynamic optimization of resources and effective control of energy consumption.This paper proposes an HPC node prediction method called iDSRformer,which introduces a multi-head sparse self-attention mechanism to improve computational efficiency and capture dependencies between multiple features,uses RMSNorm normalization to enhance stability,and adopts depthwise separable convolution to implement the feed-forward network,expanding the model's receptive field to achieve more efficient feature extraction and handle complex relationships between multiple variables.Experiments are conducted on Microsoft's Philly cluster and Alibaba's cluster-trace-gpu-v2020 datasets,respectively.The results show that compared with the currently proposed prediction models iTransformer,Transformer,patchTST,dlinear,timeMixerh,and crossformer,iDSRformer achieves average MSE reduction of 10.2%,20.2%,12.2%,19.2%,10.1%,and 13.3% and average MAE reduction of 9.9%,26.3%,10.8%,16.9%,7.4%,and 15% on the Philly task data,while on Alibaba's cluster-trace-gpu-v2020 task data,it achieves average MSE reduction of 10.2%,18.3%,16%,28.2%,15.4%,and 14.4% and average MAE reduction of 8.8%,14.4%,11.5%,19.5%,9%,and 9%,demonstrating better prediction accuracy.

Hypertension Recognition and Blood Pressure Prediction Based on Novel Hierarchical Ballistocar-diogram Signal Feature Set

ZHANG Yuchen, YE Hanyu, YAO Yuhan, JIANG Rui, YANG Gang, ZHANG Xianchao

Computer Science. 2026, 53 (7): 262-271. doi:10.11896/jsjkx.250400022

Abstract

PDF(3754KB) ( 5 )

References | Related Articles | Metrics

Accurate early identification of hypertension is crucial for preventing further progression of the disease.Ballistocardiogram(BCG) signals,which reflect the dynamic movement of the heart's center of mass due to blood flow during normal respiration and cardiac cycles,can precisely capture blood pressure(BP) variations.However,most existing BCG-based hypertension identification methods primarily rely on traditional analyses(e.g.,time-frequency or nonlinear domain techniques) to extract limi-ted features,which fail to comprehensively characterize BP-related signal patterns.In addition,most BCG-based studies focus on diagnosing hypertension rather than predicting precise BP values.To address these limitations,this paper proposes a novel multi-level BCG feature set incorporating time-frequency domain features,nonlinear domain features,fluctuation characteristics,and waveform features.Comparative analyses are conducted using classical machine learning models,such as random forest(RF),and mainstream deep learning architectures such as convolutional neural network(CNN) and deep neural network(DNN).The results demonstrate that the proposed feature set significantly improves hypertension identification:the RF model achieves an accuracy of 82.26%,and the CNN model reaches 83.57%,outperforming Liu et al.'s feature set(accuracy of 77.92% and 78.2%,respectively).For BP prediction,the DNN regression model achieves a root mean square error(RMSE) of 6.59 mmHg for diastolic blood pressure(DBP) and 3.99 mmHg for systolic blood pressure(SBP),both superior to results using Liu et al.'s feature set.Additionally,this study validates the impact of the BCG signal's segment length on the performance of BP analysis.These fin-dings highlight that the proposed feature engineering approach effectively captures BCG signal patterns critical for BP analysis,demonstrating the potential to computationally enable more convenient and accurate non-invasive,cuffless BP monitoring.

Unsupervised Dynamic Graph Change Point Detection Method Based on Variational Graph Auto-encoder

WANG Jiajun, JIAO Pengfei, ZHANG Xinxun, LI Tianpeng, GAO Mengzhou

Computer Science. 2026, 53 (7): 272-279. doi:10.11896/jsjkx.250900118

Abstract

PDF(2902KB) ( 8 )

References | Related Articles | Metrics

Detecting or identifying event-related change points in dynamic networks is becoming increasingly important,as structural variations in the network may correspond to changes in system functionality.However,many existing change point detection techniques fail to effectively capture node features.To address this limitation,this paper proposes an unsupervised dynamic graph change point detection model based on a Variational Graph Autoencoder(VGRCPD).The model integrates Variational Graph Autoencoders with recurrent neural networks and introduces a self-attention mechanism to compute the prior distribution from historical time steps.After obtaining graph embeddings,network snapshots are clustered into disjoint groups and arranged in temporal order.The resulting time series of cluster labels naturally indicates potential change points.Experimental results on multiple real-world and synthetic datasets demonstrate that the proposed method achieves superior performance,validating its effectiveness and potential in dynamic graph analysis.

Graph-structure-aware Single-cell Transcriptomic Embedding Clustering Model

YANG Hang, HUANG Ruizhang, XUE Jingjing, QIN Yongbin, CHEN Yanping, LIN Chuan

Computer Science. 2026, 53 (7): 280-288. doi:10.11896/jsjkx.250900078

Abstract

PDF(3262KB) ( 12 )

References | Related Articles | Metrics

Cell clustering,a core task in single-cell RNA sequencing(scRNA-seq) analysis,plays a crucial role in scRNA-seq data analysis.In recent years,single-cell deep embedding representation models have gained popularity due to their ability to simultaneously learn feature representation and perform clustering.However,these models still face several significant challenges,including massive data,widespread dropout events,and complex noise patterns in the transcriptome.This paper proposes a graph-structure-aware single-cell transcriptomic embedding clustering model(scGAEC).This model innovatively integrates a contrastive learning mechanism with a self-developed graph-aware approach for more in-depth embedding representation.It also designs a decoder based on the zero-inflated negative binomial(ZINB) model to reconstruct gene expression information.By employing a joint mutual supervision strategy,the model optimizes clustering loss,contrastive loss,ZINB loss,and gene expression matrix reconstruction loss in a coordinated manner,thereby enhancing clustering performance and deep learning of latent representation.Experimental results demonstrate that scGAEC achieves average performance improvements of 30.63% in NMI and 52.17% in ARI over six competing models across four single-cell RNA sequencing datasets from different sequencing platforms,significantly outperforming various state-of-the-art methods.

Single-cell Multi-omics Clustering Method Based on Missing Value Imputation and Cross-modalAlignment

ZHU Rong, HU Mengyao, DAI Lingyun, LI Feng

Computer Science. 2026, 53 (7): 289-297. doi:10.11896/jsjkx.250300157

Abstract

PDF(3729KB) ( 10 )

References | Related Articles | Metrics

Single-cell multi-omics joint clustering analysis has been extensively utilized in complex disease research.However,due to technical limitations of single-cell sequencing platforms,gene expression matrices contain substantial missing values,and inhe-rent heterogeneity across modalities pose significant challenges for clustering analysis.To solve the above problems,a single-cell multi-omics clustering method based on missing value imputation and cross-modal alignment(scCMAC) is proposed to address missing value imputation,cross-modal alignment and fusion of unpaired single-cell multi-omics data.The method uses a triple imputation strategy to dynamically localize missing values using modality-specific neighborhood structural information,enabling precise data recovery.To mitigate cross-modal heterogeneity and enhance inter-modal correlations,a bidirectional generative adversarial network is introduced for cross-modal alignment,while a graph attention network(GAT) facilitates multi-modal feature fusion.Moreover,the method performs clustering by jointly modeling cell-cell interactions and intracellular multi-omics associations.The results of comparison experiments on five single-cell multi-omics unpaired datasets show that scCMAC achieves 71.62% ACC on the PBMC dataset,which is an improvement of 2.36 percentage points over the suboptimal method.scCMAC improves the ARI to 70.48% on the SMAGE-3K dataset,which verifies its superiority in clustering tasks.

Autoregressive Sequence Reconstruction for Unsupervised Anomaly Detection in Medical Insurance

JI Wendi, WANG Yongquan

Computer Science. 2026, 53 (7): 298-307. doi:10.11896/jsjkx.260200102

Abstract

PDF(2716KB) ( 13 )

References | Related Articles | Metrics

Anomaly detection is a key technique for medical insurance fund supervision.Most existing approaches rely on complex feature engineering and domain expertise to characterize suspicious behaviors,making rules costly to build and maintain and difficult to adapt to evolving fraud patterns.Meanwhile,labels are also scarce,delayed,and noisy,further limiting the reliable deployment of supervised methods.To this end,this paper presents SeqRecon-AD(Sequence Reconstruction for Anomaly Detection),an unsupervised anomaly detection framework in medical insurance that models each account as a time-ordered sequence of reimbursed items and measures account risk by its deviation from normal transition patterns.Specifically,an autoregressive Transformer is trained with a next-item reconstruction objective to capture regular item transitions.Token-level negative log-likelihood losses are then aggregated into an account-level anomaly score via Top-k loss aggregation,which emphasizes sparse abnormal segments rather than average behavior.Experimental results on a real-world city-scale dataset show that SeqRecon-AD outperforms classical unsupervised baselines,representative sequence models as well as reconstruction-based autoencoders.SeqRecon-AD provides an effective and deployable unsupervised solution for medical insurance anomaly detection without relying on anomaly labels for training,improving AUC by 29.21% over the best unsupervised baseline.

Trustworthy IP Geolocation Method via Graph Neural Networks and Conformalized Quantile Regression

TAI Wenxin, LIU Xueting, WANG Xiaohan, ZHONG Ting, WANG Yong, ZHOU Fan

Computer Science. 2026, 53 (7): 308-314. doi:10.11896/jsjkx.250500009

Abstract

PDF(2140KB) ( 8 )

References | Related Articles | Metrics

IP geolocation,as a fundamental technology for cyberspace mapping and management,has significant applications in various domains such as network security,content recommendation,and financial risk control.With the rapid development of artificial intelligence,neural networks have become the predominant modeling paradigm for IP geolocation in recent years,with most existing studies focusing on minimizing the average geolocation error.However,in risk-sensitive scenarios,the controllability of localization errors is equally critical,and modeling paradigms that solely focus on minimizing average errors often fall short of practical requirements.To address this issue,this paper proposes a trustworthy IP geolocation method that integrates graph neural networks with conformalized quantile regression.Unlike traditional point estimation methods,the proposed method produces prediction intervals with guaranteed confidence levels,enabling verifiable control over the range of geolocation errors.Experimental results on multiple real-world datasets demonstrate that the proposed method achieves a prediction interval coverage deviation of less than 1% at the 90% confidence level,while maintaining narrow interval widths,significantly enhancing the trustworthiness and practicality of IP geolocation in risk-sensitive scenarios.

Dynamic Window Extension Mechanism for Emergency Burst Traffic Awareness in Time-sensitive Networking

LI Haoran, ZHANG Tong, ZHU Kun, REN Fengyuan

Computer Science. 2026, 53 (7): 315-323. doi:10.11896/jsjkx.250600095

Abstract

PDF(3899KB) ( 9 )

References | Related Articles | Metrics

Time-sensitive networking(TSN) is a key technology for real-time transmission,which provides low-latency and deterministic transmission guarantees for scheduled traffic through traffic scheduling and shaping mechanisms.However,in actual transmission scenarios such as autonomous driving and aviation control,the existence of emergency burst traffic poses new challenges to existing scheduling and shaping mechanisms.Traditional time-aware shapers(TAS) strictly plan time windows for scheduled traffic,making it difficult to effectively handlesuch emergency burst traffic while ensuring the transmission perfor-mance of scheduled traffic.To address this problem,this paper proposes a new dynamic window extension mechanism called E-DWE.By adjusting the queue allocation rules and expanding the scheduling window online,priority scheduling of emergency burst traffic is achieved,alleviating the transmission conflict between them and scheduled traffic.Simulation experiments under a variety of typical topologies and complex load conditions show that E-DWE can not only minimize the interference of emergency burst traffic on scheduled traffic,but also reduce the cascade delay of scheduled traffic by an average of about 45% while ensuring low-latency transmission of emergency burst traffic.In addition,E-DWE exhibits good scheduling robustness in high-concurrency emergency burst traffic scenarios,can accommodate more emergency burst traffic,and demonstrates strong engineering applicabi-lity and scheduling flexibility.

Traffic Scheduling Algorithm Based on Peak Link Load in Time-sensitive Networks

XU Jia, TANG Qun, WANG Naimin, XU Lijie

Computer Science. 2026, 53 (7): 324-335. doi:10.11896/jsjkx.250400034

Abstract

PDF(4469KB) ( 9 )

References | Related Articles | Metrics

The time-aware shaper(TAS) mechanism defined in the time-sensitive network(TSN) standard ensures bounded delay and jitter of time triggered(TT) traffic through gate control list(GCL).The GCL is generated by the scheduling algorithm and controls the opening and closing time of the TSN switch out queues.The existing scheduling methods use the shortest path to transmit TT flows,or isolate link state and routing,without considering the dynamic impact of link state on routing,resulting in excessive load on some links in the network.This paper proposes a joint routing and traffic scheduling algorithm based on peak link load(JRTSPLL).The proposed method integrates the idea of load balancing to minimize the peak load of links in the network during the calculation of the optimal path.Simulation results show that,compared with the shortest path scheduling algorithm,JRTSPLL reduces the peak link load by 29.96%,increases the scheduling success rate by 23.02%,and reduces the total transmission time by 15.14%.Compared with the equal cost multi-path scheduling algorithm,JRTSPLL reduces the peak link load by 9.52%,increases the scheduling success rate by 38.46%,and reduces the total transmission time by 14.01%.

Self-adaptive Load Balancing Strategy Based on Reinforcement Learning for SDSN

WANG Hongguang, JIANG Yiming, LIU Xiajun, BAI Luxin

Computer Science. 2026, 53 (7): 336-342. doi:10.11896/jsjkx.250500004

Abstract

PDF(3633KB) ( 14 )

References | Related Articles | Metrics

Given the challenges of satellite Internet,including limited computational resources,dynamically time-varying links,and uneven traffic distribution,this paper proposes an self-adaptive load balancing strategy based on reinforcement learning.Leveraging the control-data plane separation architecture of software-defined satellite networks(SDSN),a two-hop regional partitioning algorithm for SDSN is designed.To address the communication quality disparities between intra-orbit and inter-orbit links,link state values(μ) and weight values(w) are introduced to quantify link performance,prioritizing intra-orbit low-latency links.Built upon the Actor-Critic deep reinforcement learning framework,the SALB-RL model employs multi-agent asynchronous training to optimize key flow selection.Traffic redistribution ratios are computed via linear programming to minimize maximum link utilization while reducing end-to-end delay.A low earth orbit(LEO) Walker constellation is constructed using STK(Systems Tool Kit),a leading system engineering software,and traffic datasets derived from dynamic network topologies are used for training and validation.Experimental results demonstrate that SALB-RL achieves over 95% network-wide load balancing performance by redistributing only 10% of critical flows.Compared with state-of-the-art satellite Internet DRL models and traditional terrestrial load balancing algorithms,SALB-RL improves average load balancing performance by about 3% while ensuring more stable delay characteristics.This work highlights that SALB-RL effectively balances load balancing efficiency and routing overhead,offering an optimal solution for intelligent management of dynamic satellite networks.

Edge Load Prediction Method Based on s-TimeXer Combined Model

SHI Hongling, LI Jinhui, LI Chenghua, JIANG Xiaoping, DING Hao

Computer Science. 2026, 53 (7): 343-353. doi:10.11896/jsjkx.250300169

Abstract

PDF(3331KB) ( 9 )

References | Related Articles | Metrics

Load prediction methods in edge computing environments are crucial for the allocation and management of computing resources.Edge load data has characteristics such as volatility,noise,mutation,and time dependence.Therefore,a single prediction model is difficult to effectively extract the multi-dimensional information of load data.To address the above problems,an edge load prediction method based on thehybrid s-TimeXer model is proposed.Firstly,the FFT-SSD collaborative decomposition module is constructed.The main period of the load data is extracted by Fast Fourier Transform as the window length parameter of singular spectrum decomposition,which enhances the ability to capture periodic oscillation structure and realizes the effective separation of trend term,period term and noise term.Then,the load data is embedded as an endogenous variable,and the characteristic subsequences of singular spectrum decomposition are embedded as exogenous variables to construct a multi-dimensional feature interaction space.The time dependency of the load data is captured through the self-attention mechanism,and the dynamic interaction between the load data and the characteristic subsequences of singular spectrum decomposition is achieved through the cross-attention mechanism,thereby enhancing the contribution of the periodic component and the trend component to the prediction target.At the same time,the Hyperband Pruner algorithm is introduced to achieve efficient optimization of hyperparameters and improve prediction accuracy.Through the decomposition-embedding joint optimization architecture,while inheriting the advantages of TimeXer time series modeling,the synergistic enhancement of noise suppression and multi-dimensional information extraction is achieved.Experimental results on the ECW and Alibaba datasets demonstrate that the s-TimeXer model surpasses multiple state-of-the-art baseline methods in prediction accuracy.Specifically,on the ECW dataset,the reductions are 27.7%~63.4% for MSE and 14.3%~46.5% for MAE;on the Alibaba dataset,the reductions are 39.7%~42.5% for MSE and 18.4%~23.8% for MAE.The s-TimeXer model can effectively improve the accuracy of edge load prediction and provide strong support for resource scheduling in edge computing environments.

Multi-party Inter-satellite Collaborative Computing Offloading Algorithm for Time-varying Topologies and Dynamic Heterogeneous Resources

SHANG Kefeng, ZHANG Dan, ZHUAN Sunying, LI Dandan, LIU Yan, ZHU Kaige

Computer Science. 2026, 53 (7): 354-362. doi:10.11896/jsjkx.250500060

Abstract

PDF(3249KB) ( 12 )

References | Related Articles | Metrics

In the scenario of satellite-ground collaborative computing offloading,the satellite first computes part of the tasks and then offloads the remaining tasks to the ground for completion.This scenario typically requires the transmission of a large amount of data between the satellite and the ground,which incurs high costs for satellites with limited resources.Inter-satellite collaborative computing offloading can complete tasks without relying on ground computing devices.However,current research has not comprehensively considered the time-varying nature of satellite network topology,the dynamic and heterogeneous nature of satellite resources,which reduces the success rate of tasks.Therefore,an inter-satellite multi-party collaborative computing offloading algorithm for time-varying satellite topology and dynamic heterogeneous resources is proposed.Specifically,for the task set in each time slot,to make the offloading algorithm adapt to the time-varying satellite topology and dynamic heterogeneous satellite resources,the algorithm first collects key information such as the current time slot's satellite network topology,the resource status of each satellite,and the connection duration with the ground.Subsequently,the ground center optimizes task allocation with the dual objectives of minimizing task latency and maximizing task success rate.The primary objectives are to minimize task delay and maximize the task success rate for each task.Experimentally,dynamic topological data of the satellite network and connection time data are collected using the STK tool.The results show that,compared with the baseline algorithms,the proposed algorithm achieves a higher task success rate and lower task delay.

Visual Analysis Method for Understanding Evolution Patterns of Autonomous System Relationships

JIANG Peng, TANG Jingwei, CHEN Jiahui, PAN Xiaojie, XIA Xinyu, LIU Jian, WANG Yunchao, SUN Guodao, LIANG Ronghua

Computer Science. 2026, 53 (7): 363-371. doi:10.11896/jsjkx.251200041

Abstract

PDF(4479KB) ( 12 )

References | Related Articles | Metrics

The routing relationships of autonomous systems(ASs) exhibit high dynamics,constantly changing due to multiple influencing factors,which leads to the inefficiency and insufficient expressiveness of existing analytical methods in characterizing their structural evolution and identifying key changes.To address this,this paper proposes an interactive visual analysis approach for autonomous system topology evolution,aiming at supporting the identification and understanding of network structure changes.Firstly,a weighted network model is constructed based on AS business relationships,and core features such as structural clustering patterns,scale,influence,and connection types across different time slices are extracted through community detection.Subsequently,combining structural features with relational changes,a community-based event classification method is employed to abstract the changes in business relationships into structural events,enabling a unified representation of topology evolution.Finally,event matrices and radial views are used to perform associative analysis of community spatiotemporal evolution and their hierarchical relationships.Two case studies and user evaluation results validate the effectiveness and practicality of the proposed approach in assisting users in analyzing network topology changes.

ReGAN:Enhancing Wi-Fi Activity Recognition Under Low Packet Rates Using Image Reconstruction

MA Ruihu, HUANG Yujie, YAO Junmei

Computer Science. 2026, 53 (7): 372-380. doi:10.11896/jsjkx.250700186

Abstract

PDF(3122KB) ( 8 )

References | Related Articles | Metrics

In existing Wi-Fi sensing methods,the channel state information(CSI) under high packet rates is typically required to guarantee the sensing performance,which imposes significant burdens on communication resources and energy consumption.In the scenario of low packet rate,the loss of high-frequency dynamic information severely degrades the sensing performance.Interpolation-based recovery methods often fail to reconstruct critical features.To address these challenges,this paper proposes ReGAN,a generalized CSI reconstruction framework tailored for low-packet-rate conditions.ReGAN integrates an irregular masking strategy with a context-aggregated generative adversarial network(GAN),and employs a four-term composite loss to jointly optimize reconstruction quality at multiple levels,including pixel,stripe,spectral,and semantic.Experimental results de-monstrate that,in CSI reconstruction tasks under low packet rate conditions,the action classification accuracy of ReGAN in downstream sensing tasks is only 2～3 percentage lower than that of the original high packet rate data,while the testing accuracy of multiple shallow statistical models exceeds 83%,indicating its strong capability in structural restoration and cross model generalization under low packet rate sensing scenarios.ReGAN effectively guarantees the sensing performance under limited transmission rates without increasing communication overhead,offering a practical and efficient solution for large-scale deployment of Wi-Fi activity recognition in edge scenarios.

Survey on Security Risks and Mitigation Strategies for Generative Artificial Intelligence

CHEN Quantao, ZHANG Yangsen, WANG Pu, GUO Yalong

Computer Science. 2026, 53 (7): 381-396. doi:10.11896/jsjkx.250600196

Abstract

PDF(2522KB) ( 13 )

References | Related Articles | Metrics

With the widespread application of GAI(Generative Artificial Intelligence) technologies across multimodal domains such as image,text,and audio generation,the associated security risks and governance challenges have become increasingly prominent.This paper aims to systematically review the core security issues faced by GAI and to provide an in-depth analysis from two dimensions:the data layer and the model layer.At the data layer,five major categories of risks are examined,including privacy leakage,data tampering,unreliable data sources,bias and unfairness,and data sovereignty and compliance.At the model layer,seven major security threats are analyzed,namely adversarial examples,deepfakes,prompt attacks,backdoor attacks,model stea-ling attacks,model inversion attacks,and model jailbreak attacks.In response to these issues,this paper further summarizes recent advances in defense technologies from both academia and industry,covering differential privacy(DP),homomorphic encryption(HE),anomaly detection,access control,adversarial robustness,model watermarking,machine unlearning(MU),and knowledge editing.Finally,based on the above discussion,the paper outlines future challenges and research directions in the field of GAI security,with the aim of providing useful insights and a reference framework for the continued exploration of GAI security.

Collaborative Adversarial Training Defense Framework for Network Traffic Classification Based on Ensemble Learning and Weight Constraint

HUANG Yilu, HE Xingxing, REN Ruibin, ZENG Wenqiang

Computer Science. 2026, 53 (7): 397-405. doi:10.11896/jsjkx.250600039

Abstract

PDF(2960KB) ( 10 )

References | Related Articles | Metrics

With the growing adoption of deep learning in network traffic classification,model vulnerability to minor perturbations has become a critical security concern.Adversarial example attacks significantly compromise the reliability of real-world deployments.To address the longstanding trade-off between accuracy and robustness in conventional adversarial training,this paper proposes a collaborative adversarial defense framework that integrates ensemble learning with a dynamic weight constraint mechanism.The proposed approach assigns higher training weights to samples near decision boundaries to improve the model's sensitivity to vulnerable instances,while leveraging multi-model ensemble strategies to mitigate gradient overfitting and enhance robustness.Comprehensive experiments conducted on three widely used network traffic datasets－USTC-TFC2016,NSL-KDD,and CIC-IDS2017－demonstrate that the proposed method consistently improves robustness by over 10% across various attack types,and even exceeds 20% under high-intensity perturbations,compared with conventional baselines.The proposed framework de-monstrates exemplary scalability,affording seamless adaptation to heterogeneous network architectures and deployment contexts,thereby manifesting substantial practical utility and promising engineering applicability.

VxSymRe:VxWorks Firmware Function Symbol Recovery System Based on Binary CodeSimilarity Detection

LIU Yicong, MI Qingrong, GENG Yangyang, MA Rongkuan, JIA Yan, CAO Yan

Computer Science. 2026, 53 (7): 406-413. doi:10.11896/jsjkx.250600072

Abstract

PDF(2130KB) ( 7 )

References | Related Articles | Metrics

VxWorks firmware is widely used in critical domains such as aerospace,industrial control,and military equipment.However,to meet security and size optimization requirements,symbol tables are often stripped from released firmware versions.This significantly increases the difficulty of understanding function semantics during reverse engineering and security analysis.To address this issue,this paper presents VxSymRe－an automated function symbol recovery system specifically designed for VxWorks firmware.The system first extracts VxWorks function samples from complete ELF files and object files,and then constructs a function feature vector database that spans multiple versions,architectures,and compiler combinations.During firmware analysis,an automated preprocessing pipeline identifies the firmware version,architecture,and load base address,which enables accurate function boundary recognition.Subsequently,functions are represented via semantic-oriented graphs,and fixed-dimensional feature vectors are generated using a graph neural network.Function symbol recovery is performed through cosine similarity-based matching.Experimental results show that using both ELF and object files for database construction increases the number of function samples by 42.49% compared to using ELF files alone.In terms of preprocessing,VxSymRe is capable of correctly handling firmware from various vendors with diverse configurations.For symbol recovery,VxSymRe achieves function symbol recovery counts that are 20.86×,17.96×,and 118.2× higher than the baseline methods on three representative test samples.

Multivariate Mimic Voting Method Based on Anomaly Perception

WANG Jia, GAN Yongqiang

Computer Science. 2026, 53 (7): 414-421. doi:10.11896/jsjkx.250500059

Abstract

PDF(2744KB) ( 7 )

References | Related Articles | Metrics

In mimic defense system,the security of mimic voter directly affects the system defensive capability.Existing mimic voting algorithms typically either rely on anomaly detection to enhance the perception of error outputs of executors,or depend on heterogeneity or historical confidence to quantify the reliability of executor outputs,which leads to inexact voting output with higher-order common-mode vulnerabilities in dynamic network environments.To address above problem,this paper proposes a multivariable mimic voting based on anomaly perception.Because existing anomaly detection models always focuse on temporal or spatial information,a spatiotemporal anomaly perception model is constructed to more precisely capture the spatiotemporal cha-racteristics of executor output.Simultaneously,with the consideration of decision misjudgment caused by higher-order common-mode vulnerabilities and structural reasons of executors,higher-order heterogeneity and historical confidence are integrated with data consistency to improve the reliability of the voting results.Ultimately,an adaptive optimization strategy is desigened to adjust weight metrics to yield the optimal weighted outcome.Experimental results show that the proposed algorithm achieves an average accuracy of 98.77% on CICIDS and UNSW-NB15 datasets.Especially,a significant improvement of average about 2% over traditional algorithms on UNSW-NB15,demonstrating better stability and generalizability.

OptimalFix:Complete Framework for Efficient Detection and Patch of Vulnerabilities in SmartContracts Automatically

CHEN Shanshan, JING Ningkang

Computer Science. 2026, 53 (7): 422-432. doi:10.11896/jsjkx.250500008

Abstract

PDF(3323KB) ( 9 )

References | Related Articles | Metrics

Ethereum smart contracts have attracted significant attention in recent years,but their security vulnerabilities could result in substantial financial losses,highlighting the importance of detecting and patching these vulnerabilities.However,recent research primarily focuses on vulnerability detection rather than notable repair tools,which could result in high costs for manual repair.To address this challenge,this paper introduces OptimalFix,a novel framework that integrates vulnerability detection and patch,automatically generating secure patches for unsafe smart contracts.OptimalFix converts smart contracts into an intermediate representation and employs static detection to quickly identify and locate vulnerabilities.Subsequently,it conducts static program analysis,such as false positive filtering and program dependency analysis.The framework then generates patches using a template-based approach and selects the optimal template based on the results of the earlier static analysis,including detection outcomes and program analysis.OptimalFix is evaluated on three datasets,successfully fixing 90.2% of the vulnerabilities,with an average cost of only 900 milliseconds and a 6.7% increase in gas consumption.

Personalized Federated Learning for Concept and Label Distribution Drift

PING Fengqin, FU Xiaodong

Computer Science. 2026, 53 (7): 433-441. doi:10.11896/jsjkx.250900003

Abstract

PDF(1749KB) ( 10 )

References | Related Articles | Metrics

Personalized federated learning(PFL) combats data heterogeneity by retaining custom models for each client.Existing works primarily address label distribution drift but rarely consider non-temporal concept drift or the gradient conflicts caused by their combination,leading to slow convergence and degraded generalization in global models.This study proposes a hierarchical aggregation-dual prototype negative distillation method to enhance PFL performance in such extreme heterogeneous scenarios.The method first identifies concept-drifted clients early on through short-term training preheating,peer evaluation,and majority voting,without the need for additional private information.Then,within the trusted client set,it calculates the similarity of conditional distributions and complementarity of marginal distributions using the local model's principal components.This dynamic weighting is used in recursive hierarchical aggregation,balancing semantic consistency with feature diversity.Finally,positive and negative category prototypes are extracted from both trusted and abnormal clients to generate pseudo-samples,and a combined loss of cross-entropy and margin-based negative distillation is applied to the global model,simultaneously reinforcing correct semantics and explicitly suppressing conflicting concepts.Prototype distillation fine-tuning is also applied to abnormal clients to maintain personalized accuracy.Experiments conducted on Fashion-MNIST,CIFAR-10,and CIFAR-100 datasets in concept drift scenarios show that the proposed method achieves an average global accuracy improvement of 3.6 percentage points and the communication rounds are comparable to those of classical averaging aggregation.The research concludes that this method effectively enhances both global generalization and local adaptability in complex dual-bias environments.