Computer Science

Research on Design of Multi-level SQL Optimization Framework Based on Big Data Business Workflows

SHEN Fengyin, FAN Hongjie

Computer Science. 2026, 53 (8): 1-8. doi:10.11896/jsjkx.250900007

Abstract

PDF(2424KB) ( 152 )

References | Related Articles | Metrics

This paper proposes a multi-level SQL optimization framework tailored for business scenarios.By decoupling business logic from general-purpose computation,it establishes a collaborative architecture comprising a front-end rule engine and a back-end execution engine.The front-end SQL optimizer focuses on business-specific optimizations,while the back-end SQL execution engine is dedicated to general big data processing.By encapsulating and integrating existing unified batch-stream proces-sing engines,the framework achieves high computational efficiency.The optimization strategies are systematically divided into two levels:rule-based optimizations oriented toward business characteristics,and general execution optimizations based on the computational framework.This division enhances both modularity and iterability of the optimization strategies.The design and implementation delve into a rule-engine-driven front-end SQL optimizer,which works in conjunction with a general big data execution engine.This synergy enables business-oriented enhancements such as push-down operations for external data sources,consolidation of multiple production tasks,and effectively addresses common practical issues including data skew and code generation limits.Experiments,including external data source push-down tests,multi-task consolidation tests,and data skew load optimization tests,demonstrate performance improvements of 49.8% in field pruning and 79.75% in task merging across multiple real-world scenarios.These results indicate a substantial overall enhancement in the execution performance of big data production tasks.Furthermore,the core design philosophy of decoupling business-specific and general-purpose computation offers an innovative solution for efficient SQL optimization and execution in big data contexts.

Construction and Generation of SOR Label System for Learner Personalized Portrait

LAI Yingxu, ZHANG Yuwei, ZHUANG Junxi

Computer Science. 2026, 53 (8): 9-19. doi:10.11896/jsjkx.250600231

Abstract

PDF(3121KB) ( 117 )

References | Related Articles | Metrics

With the development of online education and personalized education,how to utilize massive online learner data to construct accurate and personalized learner portrait has become a current research hotspot.Addressing the issues in existing learner portrait research,such as neglecting individual differences and lacking systematic theoretical guidance,a method for constructing and generating a SOR label system for learner personalized portrait is proposed.Firstly,this method combines SOR theory to construct a personalized label system,extracts personalized labels,and explores individual differences among learners.Secondly,based on a multi-label classification model,it learns the complex relationship between features and personalized labels,achieving accurate prediction of personalized labels and dynamic generation of portrait labels.Finally,it constructs learner portrait based on SOR theory,revealing the relationship between environmental stimuli,individual cognition,and learning behavior.Experiment is conducted based on a public MOOC dataset,validating the effectiveness of the proposed label system through predicting the dropout rate in real teaching scenarios,achieving a 0.91% improvement in prediction accuracy over the second-best label system.The results verify the effectiveness of this method in mining learner personalized labels and constructing label systems.

Anomaly Detection in Time Series Based on Time-Frequency Contrastive Learning

LIU Yanze, HAN Bo, YUAN Jidong, SU Dongliang, REN Jia, CAI Zhiming, WANG Zhihai

Computer Science. 2026, 53 (8): 20-28. doi:10.11896/jsjkx.250700100

Abstract

PDF(5398KB) ( 124 )

References | Related Articles | Metrics

Time series anomaly detection plays a crucial role in various fields such as finance,healthcare,and industrial production.Relevant applications include detecting abnormal response codes in core banking systems within the financial sector,monitoring electroencephalographic(EEG) signals in healthcare,and providing early warnings for industrial sensor signals.In the field of multivariate time series anomaly detection,existing methods have limitations in handling complex patterns and data with sparse labels.This paper proposes a self-supervised time series anomaly detection algorithm based on time-frequency contrastive learning,which innovatively combines time-domain and frequency-domain features to build a cross-modal representation learning framework.Firstly,high-quality time-frequency consistent representations are obtained from randomly augmented samples through time-domain and frequency-domain feature extraction.Secondly,the model is optimized using a time-frequency ranking contrastive loss function and a hierarchical contrastive loss function and,which accurately control the distance between different samples,capture multi-scale information in time series,and effectively address the issue of false negative samples in contrastive learning.Finally,anomalies are identified based on sample masking and sliding window detection.Experimental results on several datasets from different fields show that the algorithm achieves state-of-the-art performance.Ablation experiments and visualization results verify the effectiveness of the proposed algorithm and its modules,providing an efficient,accurate,and robust solution for time series anomaly detection in finance,industry,and other fields.

Time Series Language Model for Continuous Glucose Monitoring Interpretation

WANG Yuqi, ZHANG Yangsen, GUO Yalong, KANG Jing, WANG Yalun

Computer Science. 2026, 53 (8): 29-39. doi:10.11896/jsjkx.260300019

Abstract

PDF(4641KB) ( 97 )

References | Related Articles | Metrics

The interpretation of continuous glucose monitoring(CGM) signals for diabetic patients is crucial for glycemic management.However,large language model(LLM) for CGM is found the deficiency of high misinterpretation rates and the generation that conflicts with actual glucose due to the poor understanding of time series.To address these issues,the time series language model for CGM interpretation(CGM-TSLM) and a CGM interpretation dataset(GLiDCGM) of “glucose series-language description” pairs are proposed.A one-dimensional convolutional neural network time-series encoder is involved to capture the key quantitative features of the CGM series,a language model is employed to encode prompt instructions and generate matching language,and an attention mechanism is introduced to align and fuse features.Finally,the transport of the CGM series to text is completed by the multimodal supervised fine-tuning method.For the GLiDCGM dataset construction,the fuzzy logic text annotation method is adopted to generate initial descriptions of curve features,and optimized into the concise and accurate medical summaries by LLM.Experiments on the GLiDCGM dataset show that,the description generation of CGM-TSLM model is superior than baseline models such as LLaMA2-7B-Chat,Qwen,T5,and BART,with average improvements of 12.22 percentage points and 14.45 percentage points on word overlap and text similarity,respectively.Experimental results prove that CGM-TSLM is able to further enhance the generation ability of CGM series summaries and provide theoretical and data support for the analysis of wearable devices physiological data.

STMVF:Novel Multi-view Graph Convolutional Network for Spatial Transcriptomics Cell Deconvolution with Dual Cross-attention Mechanism

GUO Peilin, ZOU Zhiyi, WANG Bo, LUO Jiawei

Computer Science. 2026, 53 (8): 40-49. doi:10.11896/jsjkx.250800055

Abstract

PDF(11261KB) ( 84 )

References | Related Articles | Metrics

In recent years,breakthrough advancements in spatial transcriptomics(ST) technologies have opened new avenues for understanding tissue architecture and cellular heterogeneity.However,mainstream ST technologies like 10x Visium have not yet achieved single-cell resolution,posing challenges for accurate cell type deconvolution.Although various solutions have been proposed,most methods rely solely on gene expression profiles and spatial location information,often neglecting gene expression si-milarity,which limits their performance.To address this,this paper proposes a novel computational framework,STMVF,which integrates cross-attention-based multi-view feature fusion and graph contrastive learning for cell deconvolution tasks in spatial transcriptomics.STMVF first employs a multi-view feature fusion module,which introduces a gene expression-based feature adjacency matrix to provide the model with cell state information complementary to spatial location data.Subsequently,graph contrastive learning is applied to enable the model to capture complex spatial patterns of gene expression and learn robust representations.Based on the learned representations,STMVF integrates single-cell RNA sequencing data and spatial transcriptomics data by optimizing a trainable cell-spot mapping matrix,thereby reconstructing intra-tissue cell distributions with high precision.Furthermore,to further enhance model performance,STMVF implicitly introduces spatial constraints by designing a contrastive loss function to ensure the spatial continuity of reconstructed tissue structures.Benchmarked against seven state-of-the-art methods on four simulated and real datasets,STMVF consistently demonstrates superior cell deconvolution performance.

Enhanced GNNs Across Social Networks User Identity Linkage Algorithm Based on HiddenFeatures

PAN Yuquan, YUAN Deyu, WANG Anran, JIA Yuan

Computer Science. 2026, 53 (8): 50-60. doi:10.11896/jsjkx.250500069

Abstract

PDF(3535KB) ( 83 )

References | Related Articles | Metrics

Across social networks user identification can determine whether virtual users from different social networks correspond to the same natural person.In order to address the influence of the equilibrium degree of the distribution of positive and negative samples in real datasets on the accuracy of user identity discrimination,an enhanced GNNs cross-network identity association algorithm based on hidden features is proposed.Firstly,in order to make full use of the structural information of social networks,the hidden features therein are mined,and it is clarified that nodes with larger degrees and first-order neighbor nodes play important roles.Secondly,the w-LINE algorithm is designed to generate the structural feature vectors.By concatenating them with the structural feature value vectors,the user feature vectors are obtained.Then,enhanced GNNs are proposed,including Feature-GCN,Dynamic-Weight-GAT,and GAE-VAE,for optimizing the expression of user vectors.Finally,the adaptive capsule network is introduced.With the settings of the custom routing layer and loss weights,it can flexibly cope with different positive and negative sample distributions and achieve cross-network identity association.Experiments are conducted on two real datasets.The comparative experimental results with the baseline model show that the proposed algorithm improves by more than 10% in the values of P,R,and F1.The results of the ablation experiment show that each component plays an important role in the overall performance of the proposed algorithm.

Deep Forest Models Based on CNN Feature Enhancement and Dynamic Sparse Cascading

ZHANG Xueyi, YAN Feifei

Computer Science. 2026, 53 (8): 61-70. doi:10.11896/jsjkx.260400046

Abstract

PDF(2487KB) ( 75 )

References | Related Articles | Metrics

To address the computational redundancy of multi-granularity scanning and the efficiency loss caused by feature dimension expansion in the original deep forest framework,this paper proposes a novel deep forest model－CDS-Forest－based on CNN feature enhancement and dynamic sparse cascading.The proposed model utilizes a lightweight convolutional neural network(CNN) for efficient hierarchical feature extraction,effectively replacing the computationally intensive multi-granularity scanning module.Furthermore,a dynamic sparse cascading mechanism is introduced to adaptively control model complexity and eliminate feature redundancy through selective feature integration.Experimental results on two benchmark datasets,MNIST and CIFAR-10,demonstrate the superiority of CDS-Forest.On the MNIST dataset,it achieves an accuracy of 99.14%,significantly outperforming the original gcForest(96.28%) and standard CNN(98.75%),while its training time is only 1/112 of gcForest.On the CIFAR-10 dataset,CDS-Forest achieves an accuracy of 88.76%,which is 0.48 percentage points and 4.1 percentage points higher than standard CNN and gcForest,respectively,with the training time reduced to 1/62 of gcForest and 1/3 of standard CNN.CDS-Forest model exhibits low hardware dependency and operates efficiently without high-performance GPUs,offering a cost-effective and low-deployment-cost solution.This work provides an efficient,user-friendly,and hardware-light non-neural deep learning alternative for image classification,enriching the integration of deep learning and ensemble methods.

Adaptive Frequency Tuning Approach for Spark Clusters

HE Yulin, XIAO Youqi, YANG Zhenyu, HUANG Zhexue, CUI Laizhong

Computer Science. 2026, 53 (8): 71-84. doi:10.11896/jsjkx.250800093

Abstract

PDF(6773KB) ( 69 )

References | Related Articles | Metrics

Data center energy efficiency represents a critical challenge in green computing.Distributed big data frameworks such as Spark consistently exhibit high energy consumption due to their in-memory computation architecture.Current data centers face a structural imbalance between resource supply and dynamic demand:daily workload fluctuations frequently lead to hardware idling during low-activity periods,resulting in substantial resource wastage.Moreover,uneven load distribution across cluster nodes contributes to resource fragmentation and underutilization.Existing energy optimization methods for clusters often suffer from coarse control granularity,heavy reliance on historical data,and limited adaptability to heterogeneous workloads,leading to constrained effectiveness and high operational costs.To address these limitations,this paper introduces a deep reinforcement lear-ning-based approach for cluster energy efficiency optimization.The proposed method constructs an Actor-Critic agent(PPO-Agent) using the proximal policy optimization algorithm,achieves stable convergence through policy entropy regularization,and leverages dynamic voltage and frequency scaling(DVFS) commands to precisely align hardware resources with real-time job requirements.This enables more effective reduction of overall cluster energy consumption.Experimental evaluations under the HiBench benchmark on Spark-on-YARN clusters demonstrate that PPOAgent achieves energy savings of up to 9.7% and 12.9% compared to FAESS-DVFS and GRMSSY,respectively,while consistently meeting service level agreements.Furthermore,PPOAgent reduces execution time by up to 30.3% and 49.7%,respectively,and significantly improves task completion rates under high-load conditions.These results confirm that the proposed method successfully achieves synergistic optimization of energy efficiency and performance across diverse workloads and dynamic environments,demonstrating considerable practical value.

Social Networks and Stock Price Synchronicity:Explainable Predictive Model Based on GraphAttention Network

LU Shan, LIU Yuelong, ZHAO Zhiqi, GU Jie

Computer Science. 2026, 53 (8): 85-93. doi:10.11896/jsjkx.250700115

Abstract

PDF(2778KB) ( 74 )

References | Related Articles | Metrics

Social media is increasingly becoming an important arena for financial information dissemination and investor decision-making interactions.Leveraging massive investor behavior data to analyze financial market dynamics has emerged as a frontier direction in the intersection of behavioral finance and computational social science.Information diffusion within social networks profoundly influences asset pricing efficiency.However,existing studies largely rely on structured data,such as financial factors,ma-king it difficult to capture the dynamic evolution of investors’ network structures and emotional interactions.This study focuses on the relationship between investor social networks and asset pricing efficiency.Based on large-scale social media data,it constructs an investor-following network and uses the short-text posts published by users as node features.A stock price synchroni-city prediction model is then developed using a graph attention network,and interpretability methods are introduced to identify key textual features and typical network structures that drive synchronicity increases.Empirical results show that integrating investor network structures with textual information significantly improves the accuracy of stock price synchronicity prediction.Furthermore,interpretability analyses reveal key words and subgraph patterns that exert substantial influence on synchronicity prediction.This research provides a new data-driven and interpretable analytical perspective for asset pricing and market efficiency,and extends the intersection of deep learning and network science into financial research.

Proxy-based Source-free Domain Adaptation for EEG Emotion Recognition Method

WU Shuiqing, QIU Jihao, LIU Xiang, DONG Yuxi, WEN Yimin

Computer Science. 2026, 53 (8): 94-102. doi:10.11896/jsjkx.250700141

Abstract

PDF(3187KB) ( 79 )

References | Related Articles | Metrics

Emotion recognition using electroencephalography(EEG) has made significant progress in the field of emotional brain-computer interfaces.However,traditional unsupervised domain adaptation(UDA) methods typically require simultaneous access to both source and target domain data,posing a risk of privacy leakage for source domain subjects.To estimate inter-domain differences and suppress noisy pseudo-labels without the need for source domain samples,this paper proposes a proxy-based source-free domain adaptation for EEG emotion recognition method.This method uses the row vectors of the classifier weight matrix from a pre-trained source model as class prototypes,based on which the nearest neighbor samples in the target domain are selected to construct a class-balanced proxy source domain.Subsequently,this proxy source domain is used to train the target model,and the quality of the proxy domain is improved by optimizing the class prototypes and sample selection process.Addi-tionally,the Mixup algorithm is employed to mix the features extracted from the target domain to enhance feature representation capability.A pseudo-label weighting and correction strategy is also proposed,which re-weights the classification loss through uncertainty estimation and corrects the pseudo-labels of samples with high uncertainty.Compared to mainstream source-free domain adaptation(SFDA) methods,the proposed approach achieves an average accuracy improvement of 3.24% on the SEED,SEED-IV,and SEED-V datasets,while also demonstrating significant advantages in pseudo-label quality.

Lithium-ion Battery Capacity Data Augmentation and Prediction Based on Diffusion,Denoise and Coding-Decoding Attention

LIAO Xuechao, ZOU Hang, LYU Peidong, ZENG Zhiqiang

Computer Science. 2026, 53 (8): 103-116. doi:10.11896/jsjkx.250500134

Abstract

PDF(7280KB) ( 76 )

References | Related Articles | Metrics

Lithium-ion battery capacity prediction is critical for enhancing battery economy and safety,yet existing methods face challenges of incomplete local information capture and poor data quality.To address these issues,this paper proposes a novel framework integrating seasonal-trend decomposition using Loess(STL) with advanced modeling techniques.Firstly,a diffusion model(DMnet) is developed by embedding STL into the diffusion process,where STL-extracted features serve as prior conditions to refine data generation.Secondly,an encoder-decoder attention model(STLnet) is designed by incorporating STL to better capture temporal feature dependencies.Finally,an integrated model(SDMnet) is constructed by synergizing the diffusion and prediction models,leveraging STL-derived features to dynamically regulate both data augmentation and prediction processes for high-precision capacity forecasting.Comparative experiments demonstrate that STL-guided data augmentation via diffusion effectively enhances model generalization,while the improved STLnet achieves superior prediction accuracy.Collectively,SDMnet exhibits outstanding performance in both prediction precision and generalization capability.

Zero-shot Pedestrian Trajectory Prediction Method Based on Compositional Motion

DENG Jiayan, TIAN Shirui, LIU Hou, ZHU Ningbo, DUAN Mingxing

Computer Science. 2026, 53 (8): 117-126. doi:10.11896/jsjkx.260700063

Abstract

PDF(3707KB) ( 71 )

References | Related Articles | Metrics

To address the limited explicit modeling of motion structure in existing pedestrian trajectory prediction methods and the poor generalization to unseen compositional motion caused by scarce complex combinations,this paper proposes a compositional motion zero-shot pedestrian trajectory prediction network(CZP-Net).Firstly,the motion encoding module extracts an in-teractive motion representation of the target pedestrian and neighboring pedestrians,capturing the joint influence of individual motion tendency and neighborhood interaction on future trajectories.Secondly,a motion-unit prototype bank learns reusable and composable motion representations,while a prototype dispersion constraint reduces redundancy and improves discrimination among prototypes.Thirdly,compositional motion reasoning module adaptively generates combination weights from the similarity between the interaction representation and the motion prototypes,producing a compositional representation for unseen motion patterns.Finally,a shared motion decoder predicts multimodal future trajectories.The future is modeled as a conditional distribution over combinations of motion units,and the combination weights are optimized under the maximum-entropy principle to provide probabilistic and semantic interpretability.On the zero-shot compositional motion task,CZP-Net reduces average displacement error(ADE) and final displacement error(FDE) by 47.46% and 30.47%,respectively,compared with the strongest baseline.On the long-tailed compositional motion task,ADE and FDE are reduced by 53.41% and 45.08%,respectively.

Sparse-view Gaussian Splatting Consistent Reconstruction with Mask-guided Generative Prior

GUO Chenhui, CHEN Zebin, TAN Guang

Computer Science. 2026, 53 (8): 127-138. doi:10.11896/jsjkx.260200106

Abstract

PDF(6957KB) ( 65 )

References | Related Articles | Metrics

Gaussian splatting achieves a favorable balance between efficiency and reconstruction quality.However,under sparse-view conditions,insufficient observations lead to under-constrained geometry and appearance,resulting in structural incompleteness.Consequently,novel view synthesis often suffers from missing content in unobserved regions and rendering artifacts,degra-ding visual realism and multi-view consistency.To address these issues,this paper proposes a generative prior-guided optimization framework for sparse-view Gaussian splatting reconstruction.Specifically,this paper builds a mask-guided restoration diffusion(MRD) model based on the pretrained degradation removal diffusion model DiFix,and further fine-tunes it on establishedtraining dataset.MRD introduces a lightweight mask prediction branch(Mask Head) to localize restoration regions primarily correspon-ding to missing areas,while also responding to certain rendering artifacts.During training,a mask-guided selective noise injection strategy is adopted,where random noise is mainly applied to the missing regions to enhance completion performance,while preserving the ability to correct rendering artifacts.The generated results from MRD are then periodically incorporated into the Gaussian splatting training as pseudo-observations,and a progressive fusion strategy is employed to gradually expand the viewpoint span of pseudo-observations in a controlled manner,thereby strengthening constraints in unobserved regions and improving multi-view consistency.In addition,a region-weighted constraint is introduced to reduce the impact of uncertainty in pseudo-observations on the optimization process.Experiments on the DL3DV and Mip-NeRF 360 datasets demonstrate that the proposed method is compatible with both 2DGS and 3DGS pipelines and achieves superior performance across multiple evaluation metrics.Qualitative results further show that the proposed method produces more complete and stable reconstructions under sparse-view settings,significantly improving visual quality and multi-view consistency in novel view synthesis.

Improved DETR-based Method for CAD Graphic Element Recognition

ZHANG Zhe, LIU Junjie, ZHANG Baili

Computer Science. 2026, 53 (8): 139-147. doi:10.11896/jsjkx.251100082

Abstract

PDF(4784KB) ( 81 )

References | Related Articles | Metrics

Graphic element recognition constitutes a critical step in the intelligent analysis and compliance review of architectural drawings.Although the DETR(Detection Transformer) framework demonstrates significant potential in general object detection,notable challenges persist when applied to complex CAD graphic element recognition tasks.Specifically,DETR's limitations primarily stem from two aspects:1) Its feature extraction mechanism exhibits insufficient adaptability to scale variations among graphic elements,failing to simultaneously capture fine details of small components and large-scale structural information effectively,which results in missed detections of small elements and ambiguous boundary recognition for large ones;2) The sparse spatial distribution of informative pixels within graphic elements hinders efficient attention focusing on key regions,making the mechanism susceptible to substantial background interference.To address these issues,three key improvements to the DETR framework are proposed.A Swin Transformer based multi-scale feature extraction module is adopted to enhance robust representational capability for elements of varying sizes.A spatially sparse sampling attention mechanism is introduced to optimize the sampling efficiency of informative features and improve perception performance in critical areas.A contrastive denoising training strategy is integrated to strengthen model localization capability for sparse informative features.Experimental results on a real-world architectural dataset demonstrate that the proposed method achieves a mAP_[0.5-0.95] of 0.675,significantly outperforming existing mainstream object detection models.This validates the proposed method's effectiveness and practical engineering value for architectural graphic element recognition task.

Zero-shot Skeleton-based Action Recognition Based on Class Knowledge Fusion and Latent SpaceOptimization

ZHOU Lijuan, LIU Zhihuan, LI Xinran, NIU Changyong

Computer Science. 2026, 53 (8): 148-155. doi:10.11896/jsjkx.250600091

Abstract

PDF(2349KB) ( 74 )

References | Related Articles | Metrics

Considering the importance of different multi-source class knowledge and the independence of various body movements,this paper proposes a zero-shot skeleton-based action recognition method based on category knowledge fusion and latent space optimization.The method integrates multi-source category knowledge through a sparse attention mechanism with importance sampling,and introduces a total correlation loss to reduce redundant correlations among independent variables in the latent space.Specifically,it firstly employs pre-trained models to extract skeleton features and multi-source textual features,then learns text fusion representations through an importance sampling-based feature fusion approach.Subsequently,a total correlation-constrained generative cross-modal alignment method establishes semantic associations between skeleton and textual features.Fina-lly,latent space features generated from unseen class representations are utilized to train classifiers for recognizing unseen actions.Experiments on NTU RGB+D,NTU RGB+D 120,and PKU-MMD datasets demonstrate that the proposed method significantly outperforms all existing mainstream approaches.

Handwriting Gender Recognition Method Based on Multi-scale Directional Attention Transformer

CAI Yi, WANG Xiaobin, CHEN Ruili, XU Jinfeng

Computer Science. 2026, 53 (8): 156-164. doi:10.11896/jsjkx.250600092

Abstract

PDF(3784KB) ( 71 )

References | Related Articles | Metrics

Handwriting gender recognition,as an important branch of biometric recognition,has broad application prospects in identity verification and criminal investigation.Traditional methods mainly rely on manual feature extraction,which has limitations such as incomplete feature representation and poor cross-language adaptability.To address these problems,a multi-scale directional attention Transformer for handwriting gender recognition(MSDAttFormer) is proposed.This method innovatively combines the global modeling capability of Transformer with the directional features unique to handwriting,integrates handwriting features at different levels through a multi-scale feature fusion module,introduces a handwriting directional attention module to accurately capture gender-related directional differences,and employs an improved Transformer encoder to establish long-range dependencies between features.A Chinese handwriting gender recognition dataset(CHAP) containing 1 620 writers is constructed,and comprehensive experimental validation is conducted on both CHAP and Hebrew handwriting dataset(HHD).The results show that MSDAttFormer achieves an accuracy of 84.46%,F1-Score of 83.97%,and ROC AUC of 90.25% on the CHAP dataset,and an accuracy of 85.71% on the HHD dataset,significantly outperforming existing methods and validating the effectiveness and cross-language generalization capability of the proposed method.

Feature Aggregation with Joint Tracking:Video Object Detection in Occlusion Scenarios

WANG Jiahui, WANG Hongyu, HAO Yingguang

Computer Science. 2026, 53 (8): 165-173. doi:10.11896/jsjkx.250700148

Abstract

PDF(3671KB) ( 72 )

References | Related Articles | Metrics

Video object detection(VID) aims to exploit temporal information between video frames to assist object detection.Unlike still-image detection,VID can leverage rich temporal context by referencing features from other frames to enhance the detection performance.Therefore,existing methods generally focus on aggregating features across multiple frames to construct temporally enhanced feature representations and improve detection accuracy.However,in video sequences,objects often undergo appearance degradation due to factors such as occlusion and motion blur,making it difficult for single-frame detectors to generate high-quality candidate regions.This,in turn,can cause the subsequent cross-frame feature aggregation to fail.To address this issue,this paper proposes a tracking-assisted video object detection method designed to mitigate aggregation failures caused by missing candidate regions.Specifically,it introduces a lightweight tracker and designs a feature selection module to jointly utilize detection outputs and motion priors provided by the tracker.It then employs the WBF(Weighted Box Fusion) algorithm to merge detection boxes with tracking predictions,producing more reliable candidate regions under challenging conditions such as occlusion.Finally,it extracts these high-quality candidate regions and their corresponding trajectory features for subsequent temporal aggregation.Experiments on the ImageNet VID dataset demonstrate that the proposed method effectively enhances robustness in occluded scenarios and achieves higher detection accuracy compared with previous VID approaches.

Hierarchical Lightweight Micro-expression Recognition Based on Optical Flow Partitioned FeatureFusion

HU Changyu, FAN Xinyu, ZHANG Zhi, DING Zixu, ZHANG Zhengyue, PENG Juhong

Computer Science. 2026, 53 (8): 174-181. doi:10.11896/jsjkx.250500012

Abstract

PDF(2664KB) ( 68 )

References | Related Articles | Metrics

Micro-expression recognition remains a challenging task due to the difficulty in capturing subtle facial movements,li-mited dataset sizes,and high computational costs.To address these issues,this paper proposes a lightweight and efficient model,named HLFM-OFP(Hierarchical Lightweight Model with Feature Fusion Based on Optical Flow Partitioning).The model employs TV-L1 optical flow for motion extraction and utilizes a hierarchical lightweight architecture to enrich the representation of low-level features.STModel(Spatiotemporal Feature Fusion Module)is further introduced to jointly capture spatial and temporal dynamics.Experimental results on composite datasets demonstrate that HLFM-OFP contains only 2.8 million parameters and reduces training time by 88.45%,while achieving 94.23% accuracy,91.23% UF1,and 91.66% UAR.Experiments show that the proposed model reduces parameters while improving accuracy,and demonstrates strong robustness on the more challenging SMIC and SAMM datasets.

Localization of Diffusion-based Image Editing via Inversion-Reconstruction

LING Yun, LI Haodong

Computer Science. 2026, 53 (8): 182-190. doi:10.11896/jsjkx.251200069

Abstract

PDF(4324KB) ( 67 )

References | Related Articles | Metrics

Image tampering localization aims to detect and locate regions in an image that have been edited or forged.With the rapid proliferation of diffusion models for image generation and editing,even ordinary users can locally modify images with high photorealism via simple textual prompts,posing new challenges to existing forensic techniques.To tackle diffusion-based image editing,this paper proposes a tampering localization method built upon DDIM inversion and multi-step reconstruction consistency.The key observation is that when an edited image undergoes DDIM inversion followed by reconstruction along the denoising trajectory,edited(diffusion-generated) regions can be stably recovered,whereas pristine,unedited regions present noticeable inconsistencies due to their mismatch with the diffusion prior.Leveraging this insight,this paper designs a dual-stream network in which one branch processes the original image and the other one handles a series of reconstructed images obtained at multiple diffusion timesteps.A self-attention mechanism is introduced to explicitly model the dynamic reconstruction behavior across timesteps,while cross-attention is used to fuse domain-specific features from the pristine and diffusion-edited regions.To mitigate data scarcity,it also constructs a diffusion-edited image dataset with ground-truth masks by using several mainstream diffusion models.Experimental results on multiple benchmarks show that the proposed me-thod substantially outperforms existing approaches on both seen and unseen types of images,demonstrating strong robustness and generalization.

Teacher Trajectory Recognition and Reconstruction in Smart Classrooms via Multi-source Fusion

WANG Yifan, YANG Peixuan, LU Yuansuo, LIU Mengjun

Computer Science. 2026, 53 (8): 191-200. doi:10.11896/jsjkx.251200195

Abstract

PDF(3289KB) ( 73 )

References | Related Articles | Metrics

Teacher trajectory reconstruction is fundamental for modeling the spatio-temporal characteristics of teaching activities in video-based classroom behavior analysis.However,real classrooms involve frequent occlusions,viewpoint variations,and complex interactions,which make single-camera tracking methods struggle to maintain trajectory continuity and stability.To address these challenges,this paper proposes a multi-camera teacher trajectory tracking and reconstruction approach for smart classroom scenarios.By integrating multi-source observations,unified ground-plane mapping,and cross-view constraints,the proposed methodenables robust cross-view association and continuous reconstruction of multi-view trajectories.Moreover,it establishes an eva-luation framework focusing on trajectory continuity,interruption frequency,and motion smoothness to comparatively analyze different methods under occlusion and viewpoint transition conditions.Experimental results show that the proposed approach effectively reduces trajectory interruptions in complex dynamic scenes and achieves better trajectory continuity and motion smoothness,demonstrating its effectiveness and robustness for teacher trajectory reconstruction in real smart classroom environments.

Improved YOLOv11n Model for Small Target Detection in UAV Aerial Images

WANG Jingyang, XUE Weimin, HUANG Min, WU Shaoguang

Computer Science. 2026, 53 (8): 201-208. doi:10.11896/jsjkx.250800100

Abstract

PDF(3712KB) ( 86 )

References | Related Articles | Metrics

With the continuous innovation of target detection algorithms and UAV(Unmanned Aerial Vehicle) technology,detecting small targets in UAV aerial images has become a hot research topic.UAV aerial images have problems such as a high proportion of small targets,dense distribution,and complex environments.To address these issues,this paper proposes an improved UAV aerial image small target detection model MPI-YOLO based on YOLOv11n.Firstly,the MSFAM(Multi-Scale Feature Aggregation Module) replaces the first two C3k2 modules in the Backbone of YOLOv11n to enhance the feature extraction ability for small target information.Secondly,to solve the problem of a high proportion of small targets in UAV images,a new P2 small target detection layer is added,so that small targets can be better detected.Finally,IBDFPN(Improved Bidirectional Dense Feature Pyramid Network) is proposed,which extends the level range of the pyramid network and introduces skip connections to enhance the cross-layer interactions between shallow and deep features,achieving efficient multi-scale information fusion.The VisDrone2019 dataset is utilized for ablation experiments and comparative experiments.The ablation experiments verify the effectiveness of each improved strategy.The comparative experimental results show that MPI-YOLO outperforms other comparison mo-dels in small target detection accuracy.Compared with the benchmark model YOLOv11n,the mAP50 of MPI-YOLO increases by 9.1 percentage points.Generalization experiments are conducted on the UAVDT dataset and Tinyperson dataset,verifying its good generalization ability.

Lightweight Low-resolution Face Recognition via Hierarchical Dynamic Feature Generation Distillation

ZHONG Rui, YAN Hongwei, LIU Jiawei

Computer Science. 2026, 53 (8): 209-218. doi:10.11896/jsjkx.250500082

Abstract

PDF(3049KB) ( 69 )

References | Related Articles | Metrics

For low-resolution face recognition scenarios demanding high accuracy,efficiency,and lightweight design,this study proposes a HDFGD(Hierarchical Dynamic Feature Generation Distillation) framework.Based on feature redundancy theory and knowledge distillation,the framework constructs a dynamic hierarchical feature generation module to enhance the feature representation capabilities of lightweight student networks.The module expands feature diversity through channel-wise convolutions,employs channel attention mechanisms to focus on critical semantic information,and integrates adaptive compression mechanisms to reduce computational complexity,effectively mitigating semantic gaps across resolution-specific features.Simultaneously,the ArcFace loss function is incorporated to optimize angular-margin constraints in the feature space,thereby strengthening intra-class compactness and inter-class separability,and ultimately improving the accuracy of low-resolution face recognition.Experimental results demonstrate that the HDFGD framework achieves a maximum 2.1% improvement in accuracy over mainstream methods on the CIFAR-100 classification task.It also attains recognition accuracies of 95.87% and 97.02% respectively on two benchmark low-resolution face re-cognition datasets,significantly outperforming existing lightweight face recognition algorithms.Ablation studies confirm that the combination of dynamic feature generation with hierarchical alignment strategies and the ArcFace discriminative supervision mechanism simultaneously achieves triple breakthroughs:model lightweighting,cross-resolution feature robustness,and recognition accuracy enhancement.This framework provides an efficient technical solution for real-time face recognition tasks in resource-constrained scenarios.

Survey on Mutually Augmenting Technologies and Applications of Large Models and KnowledgeGraphs

JIA Zishuo, ZHANG Jian’ge, HE Haofeng, FENG Shizhong, LIU Yilin

Computer Science. 2026, 53 (8): 219-228. doi:10.11896/jsjkx.250700129

Abstract

PDF(2005KB) ( 79 )

References | Related Articles | Metrics

With the development of artificial intelligence,the capabilities of large language models(LLMs) and knowledge graphs(KGs) in the field of natural language processing gained a lot of attention.Endowed with robust natural language understanding and generation capacities,LLMs have exhibited prominent emergent abilities in tasks including open-domain question answering and text generation.Nevertheless,LLMs are confronted with challenges such as inadequate interpretability and the existence of hallucinations in knowledge representation.Similarly,knowledge graphs provide interpretable symbolic support for complex reasoning and decision-making through structured knowledge representation,yet suffer from high construction costs and incomplete content.Therefore,the mutual enhancement technology of large language models and knowledge graphs has become a key research direction.This paper systematically reviews the relevant knowledge of large language models and knowledge graphs and introduces the mutual enhancement technologies between large language models and knowledge graphs,including knowledge graph-enhanced large models,large language model-enhanced knowledge graphs,and the mutual enhancement and collaboration between large language models and knowledge graphs.Additionally,it elaborates on the application of this synergistic framework.Finally,it summarizes the challenges and prospects of the mutually reinforced large language model-knowledge graph system,providing references for subsequent studies of this technology.

Review of Music Artificial Intelligence Driven by Large Language Models

LIU Jing

Computer Science. 2026, 53 (8): 229-244. doi:10.11896/jsjkx.260700167

Abstract

PDF(5473KB) ( 84 )

References | Related Articles | Metrics

Large language models and other emerging AI technologies are reshaping music AI from generation-oriented demonstrations toward systematic research on music understanding,interactive creation,and scenario-based applications.This review surveys music AI in the era of large models,covering multimodal understanding,human-AI collaborative creation,personalized and functional applications,industrial ecosystems,and copyright and ethical issues.Firstly,it examines how LLMs are integrated with symbolic music,audio representations,and cross-modal alignment,highlighting the persistent gap between symbolic reasoning and audio perception.Then,it summarizes interactive creation systems in terms of intention expression,control allocation,and human preference alignment,and discusses differences between professional and general users.Applications in conversational re-commendation,music therapy,educational support,and film scoring are reviewed,with emphasis on the limitations caused by inconsistent evaluation criteria and insufficient cross-scenario validation.Finally,AI music copyright ownership,training data tra-ceability,the boundaries of human creativity,and the risks to cultural diversity are discussed.Furthermore,future directions are identified,including audio-native music foundation models,music agents,long-range structural modeling,closed-loop evaluation,copyright governance,and cross-cultural music AI,providing reference for understanding the technological progress,application boundaries,and governance issues of music artificial intelligence.

Leveraging Multi-source Contextual Knowledge-enhanced Graph for Traffic Forecasting

QUAN Jingtao, ZHANG Lei, LIU Bailong, WANG Feifan

Computer Science. 2026, 53 (8): 245-256. doi:10.11896/jsjkx.250900014

Abstract

PDF(2702KB) ( 70 )

References | Related Articles | Metrics

Traffic prediction aims to forecast future traffic conditions based on historical traffic conditions and is a crucial research topic in the field of intelligent transportation systems.Although existing traffic prediction models have achieved considerable progress in modeling complex spatio-temporal patterns,they still heavily rely on advanced deep learning techniques and exhibit limitations in representing multi-source contextual knowledge and integrating global graph-topological information.Consequently,they fail to fully capture the diverse context about external environments and internal networks,which constrains their perfor-mance.To address these issues,this paper proposes MCK-GWN(Multi-source Contextual Knowledge-enhanced Graph WaveNet),a graph-based traffic prediction model enhanced by multi-source contextual knowledgegraph.In terms of multi-source knowledge representation,a multi-source contextual knowledge graph is constructed to enrich the POI(Point of Interest) context of traffic nodes and the semantic context between nodes,effectively characterizing both external and internal contextual knowledge.In terms of multi-source knowledge fusion,the POI heterogeneity-aware unit focuses on modeling external environmental attributes,while the semantic path-aware unit focuses on internal network semantic relations,thereby enhancing the integration capability of global graph-topological information.Experimental results on the SZ-TAXI dataset demonstrate that MCK-GWN achieves the best performance across all evaluation metrics,reducing the mean absolute error(MAE) by 3.15% compared with the state-of-the-art baseline model KMHNet.

Fake News Detection Model Based on Cross-modal Feature Fusion and Alignment

YANG Chenguang, LU Jicang, GUO Jiaxing

Computer Science. 2026, 53 (8): 257-265. doi:10.11896/jsjkx.250700054

Abstract

PDF(2080KB) ( 70 )

References | Related Articles | Metrics

Fake news is usually spread by exaggerated,distorted or misleading statements,which can shape negative social opinion and seriously endanger public safety.The current fake news is usually multi-modal,and the existing detection methods often fuse the features of each modality after extracting them separately,ignoring the correlation between modalities,which makes it difficult to fully capture the details and correlation information,resulting in unsatisfactory detection performance.To solve these pro-blems,this paper proposes a detection model based on cross-modal feature fusion and alignment(CMFFA).CMFFA optimizes the modes of feature extraction,fusion and classification,extracts modal features from macro and micro perspectives,enhances features through the attention mechanism,and evaluates the ambiguity between modalities by calculating the similarity between modalities,so as to adaptively perform cross-modal feature fusion.Firstly,the pre-trained model is used to encode the single-modal features and cross-modal features of text and image.Then,in the cross-modal feature fusion,the proportion of cross-modal features is adaptively adjusted through the ambiguity analysis between modalities,so as to better realize the fusion of cross-modal features and improve the performance of fake news detection.Experimental results show that,compared with the existing fake news detection methods on the public Chinese and English fake news datasets,the proposed model significantly improves the F1 value,precision and recall rate,verifying its effectiveness and superiority.

From Bytes to Semantics:New Paradigm for Tibetan Named Entity Recognition

LI Qingkai, QUN Nuo, NI Shengqiao, YANG Jin

Computer Science. 2026, 53 (8): 266-275. doi:10.11896/jsjkx.260300033

Abstract

PDF(3213KB) ( 82 )

References | Related Articles | Metrics

Tibetan named entity recognition faces several practical challenges,including scarce annotated data,unreliable tokenization,and degraded performance on long entities.To address these issues,this paper proposes a NER framework based on byte-level global modeling and a gated feature fusion mechanism.Specifically,the framework employs the Byte Latent Transformer to directly model raw UTF-8 byte sequences,and leverages its entropy-based dynamic patching mechanism to obtain global contextual representations without relying on a fixed vocabulary.This effectively avoids entity boundary corruption introduced by conventional tokenizers and improves robustness to out-of-vocabulary patterns.Meanwhile,a gated linear conditional fusion module is designed to inject aligned global contextual features into local boundary-preserving representations with controllable strength in an adaptive,position-wise manner,thereby forming complementary features and enhancing span modeling capability.On this basis,the model combines BiLSTM and CRF to perform sequence modeling and label decoding,jointly preserving global semantic modeling and local boundary discrimination.Experimental results on the TibetanAI_NER and TibNER datasets show that the proposed model improves the F1 score by 11.5 and 4.53 percentage points,respectively,over the corresponding state-of-the-art baselines.Ablation studies further verify the synergistic contributions of the global features and the gated fusion mechanism.Overall,the proposed framework consistently strengthens entity boundary discrimination under segmentation-constrained and low-resource scenarios,with particularly clear gains on long entity recognition,providing a feasible modeling approach for NER in low-resource languages.

CCSFR:Collaborative-Content Semantic Fusion for Review-enhanced Recommendation

ZHOU Haobin, LU Yunhao, QIN Jun, JIAO Xintao, ZENG Biqing

Computer Science. 2026, 53 (8): 276-284. doi:10.11896/jsjkx.250800027

Abstract

PDF(3142KB) ( 62 )

References | Related Articles | Metrics

Review-based hybrid recommendation systems explore the semantic aspects of user preferences by integrating user reviews into collaborative filtering methods.However,these methods ignore the semantic gap between the semantic reviews and user interactions,and thus insufficiently exploit the inherent similarities among them.On the other hand,most existing hybrid re-commendations deeply couple collaborative signals with semantic reviews,which fails to model recommendations in a fine-grained manner.To overcome these limitations,this paper proposes a collaborative-content semantic fusion for review-enhanced recommendation(CCSFR) model.The model leverages the similarity between collaborative signals and review semantics to learn more comprehensive user interests and item features,and to enhance the embedded representations of users and items.To fully leverage the similarities between collaborative signals and review semantics,a hybrid content representation learning module is designed to decouple collaborative filtering and content-based recommendation,modelling user interactions and review semantics as interaction graphs and aspect graphs,respectively.Subsequently,a graph convolutional neural network is utilized to capture both collaborative signals and semantic features to learn embedded representations of users and items in both collaborative and content modes.Building on this,a collaborative semantic fusion strategy is employed to obtain unified and high-quality representations by integrating heterogeneous features.This strategy unifies the collaborative and content modes in a shared semantic space through semantic alignment,thereby reducing the semantic gap between them.Furthermore,graph contrastive learning is adopted to capture the similarity between the two modes by optimizing the mutual information between both modes and promoting complementary integration to enhance the quality of the representation.Experiments on three datasets show that CCSFR improves average 8.5% and 5.8% on nDCG compared to the optimal baseline,and 8.2% and 4.6% on recall metrics.The results of these experiments fully demonstrate that the collaborative semantic fusion strategy can effectively integrate collaborative signals and review semantics to generate more discriminative personalized representations,thereby improving recommendation performance.

Agentic Retrieval Augmented Generation Framework Based on Retrieval Task Planning and Reflection Mechanism

ZHANG Haoran, HAO Wenning, JIN Dawei, CHENG Kai, LIU Junyang

Computer Science. 2026, 53 (8): 285-297. doi:10.11896/jsjkx.250600116

Abstract

PDF(3913KB) ( 101 )

References | Related Articles | Metrics

When large language models(LLMs) handle knowledge-intensive tasks,issues of hallucinations and knowledge timeliness restrict their application in high-reliability fields.Traditional RAG(Retrieval-Augmented Generation) systems help LLMs expand their knowledge boundaries by invoking external knowledge;however,they suffer from limitations such as insufficient semantic relevance and rigid retrieval strategies in multi-hop reasoning tasks.To break through the aforementioned bottlenecks,this study proposes an intelligent iterative retrieval-augmented generation framework(PR-RAG) based on retrieval task planning and a verification-reflection mechanism.Specifically,this framework employs a task planning module to conduct hierarchical classification of questions and decomposition of sub-problems,acquires contextual information via an adaptive hybrid retrieval strategy that integrates vector retrieval and graph retrieval,and ultimately realizes closed-loop optimization through a dual verification mechanism(including logical verification and factual verification).Experimental results show that the exact matching of PR-RAG on datasets such as HotpotQA and 2WikiMQA are increased by an average of 12.8% compared with baseline methods,which effectively improves the accuracy and robustness in handling complex knowledge-intensive tasks.This study confirms that the synergistic feedback between hierarchical retrieval task planning and the verification-reflection mechanism can significantly enhance the knowledge reasoning capabilities of LLMs and the reliability of the system.

LSQ-RAG:Retrieval-enhanced Generation Framework Based on LLM-enhanced Ranker

ZHANG Zhonglin, XIA Hang

Computer Science. 2026, 53 (8): 298-306. doi:10.11896/jsjkx.250600179

Abstract

PDF(2916KB) ( 89 )

References | Related Articles | Metrics

Large language model(LLM) has demonstrated remarkable generation and reasoning capabilities in terms of conversation,reasoning,and knowledge retention,but they still suffer from numerous limitations,such as generating answers containing hallucinations,relying on outdated parameterized knowledge,and poor model interpretation.RAG(Retrieval Augmented Generation) addresses these issues by integrating non-parametric data stores.However,directly integrating information retrieval or training these components end-to-end often leads to suboptimal results or computational inefficiencies.To this end,this paper proposes the LSQ-RAG framework,which enhances the contextual understanding capabilities of the LLM with an LLM-supervised ranker and improves the quality and accuracy of the provided passages.LSQ-RAG fine-tunes the LLM to follow instructions and discriminatively use the provided information.Subsequently,it leverages the fine-tuned LLM to generate ranking scores,which serve as a supervisory signal for training the ranker.By leveraging the power of the LLM,the proposed approach eliminates the reliance on manual annotations in ranker training while achieving higher performance.Experimental results demonstrate that LSQ-RAG outperforms existing retrieval-augmented LLMs on open-domain QA and fact-checking tasks,while exhibiting consis-tent performance improvements when applied to different LLMs,highlighting its versatility and effectiveness.

Optimization in Cross-field of Manufacturing and Transportation by Combining Reinforcement Learning and Artificial Hummingbird Algorithm

LIAN Zhaoyang, SI Bailu

Computer Science. 2026, 53 (8): 307-315. doi:10.11896/jsjkx.250700180

Abstract

PDF(3505KB) ( 78 )

References | Related Articles | Metrics

Although the research on swarm intelligence optimization algorithms inspired by animal and human behaviors has made good progress in function optimization,further exploration is needed in the research of cross-field general algorithms.This paper proposes a cross-field optimization algorithm combined with reinforcement learning and artificial hummingbird algorithm(AHA).In the guided foraging and territorial foraging processes of AHA,hummingbirds and different foraging behaviors are used as agents and actions to build the reinforcement learning architecture.By using a reinforcement learning reward mechanism,the hummingbird agent selects appropriate actions to optimize the movement range of individual hummingbirds,ultimately improving the optimization effect of the algorithm.In order to test the optimization effect and universality of the algorithm,comparative experiments are conducted on function optimization problems,flexible workshop scheduling(FWS) problem,location optimization problem of logistics center(LLC),and routing optimization problem of unmanned aerial vehicle(UAVRO) in oil factories.Experimental results demonstrate that the individuals obtained after the same number of iterations of the proposed algorithm are superior,and the solutions obtained in the corresponding field are relatively better.

Element-aware Screening Method for Popularization Cases

MAO Yixiao, WANG Zixiao, ZHANG Baili, ZONG Shaohao

Computer Science. 2026, 53 (8): 316-325. doi:10.11896/jsjkx.260500095

Abstract

PDF(2645KB) ( 73 )

References | Related Articles | Metrics

The screening of legal popularization cases is a critical mission in building a smart justice system.Traditional manual screening is costly and inefficient,while existing text classification methods struggle to accurately discern fine-grained elements related to educational value,such as key facts,judgment results,and social hot spots,thus limiting identification accuracy.To address this challenge,a legal popularization case dataset is constructed and an element-aware pre-trained model is proposed.Speci-fically,the dataset is built via cross-platform data collection and regular expression matching,clarifying the structured distribution of elements and partitioning them into three complementary semantic spaces:fact view,judgment view,and label view.A fact encoder exploring core disputes,a judgment encoder extracting warning features,and a label encoder capturing hierarchical dependen-cies are respectively utilized to extract deep features.Moreover,a multi-view feature fusion module is developed,employing independent routing and shared experts to achieve dynamic fusion of key elements across different views.Experimental results on the self-built dataset demonstrate that the proposed model effectively improves the accuracy and robustness of legal educational value identification,providing a feasible auxiliary solution for automated screening.

Gated Bidirectional Mamba Multimodal Feature Fusion Framework for Drug-Target InteractionPrediction

REN Yanzhang, GAO Tai, LI Ying, WANG Bin

Computer Science. 2026, 53 (8): 326-335. doi:10.11896/jsjkx.250500096

Abstract

PDF(3568KB) ( 69 )

References | Related Articles | Metrics

Drug-target interaction(DTI) prediction is a core component of drug discovery and repurposing.Existing models face significant challenges in multiscale modeling of target sequences and multimodal feature fusion:traditional methods based on local convolutions lose global dependencies,while Transformers suffer from quadratic complexity for long sequences,and heterogeneous feature fusion often triggers semantic conflicts and overfitting.To address these issues,this paper proposes G2MambaDTI,a novel framework based on a gated bidirectional Mamba network for multimodal feature collaborative modeling.The method employs a cascaded CNN and Transformer encoder architecture to process target sequences,utilizing an adaptive gating mechanism to balance local functional motifs and global dependency features.Furthermore,a cross-modal feature calibration module with adaptive gates is introduced to enhance critical interaction pattern representations.Finally,a bidirectional Mamba architecture is integrated to capture cross-modal long-range interactions between drug molecular graphs and target sequences,leveraging its selective state-space modeling and linear complexity to significantly improve efficiency for long sequences.Experimental comparisons on four public datasets against five other deep learning models demonstrate that the proposed method outperforms existing approaches across all performance metrics,validating its superiority in DTI prediction.

Automated Judicial Document Generation for Criminal Cases Based on Multi-agent Collaboration

HAN Linrui, SONG Gaojie, ZHENG Ri, LI Bing, CUI Yan

Computer Science. 2026, 53 (8): 336-356. doi:10.11896/jsjkx.251100160

Abstract

PDF(4772KB) ( 90 )

References | Related Articles | Metrics

To address the challenges of automated judicial document generation for criminal cases,including limited accuracy in sentencing prediction,limited adaptability to complex case circumstances,and insufficient compliance with formal writing stan-dards,this paper proposes a multi-agent collaborative automated generation method for judicial documents(MAC-AG).Following an “understanding-planning-execution-generation” pipeline,MAC-AG is built on a collaborative framework composed of four specialized agents.The case fact analysis agent,powered by a general-purpose large language model(LLM),is responsible for factual determination tasks such as case element extraction,dispute focus identification,and case type classification.The case classification agent incorporates retrieval-augmented generation(RAG) to retrieve relevant legal provisions and similar cases,thereby supporting charge determination and simple-complex case routing.The judgment prediction agent employs a deep learning regression model to calibrate the LLM’s preliminary judgment predictions,improving the accuracy of sentence term and fine prediction.The document generation agent,built on a domain-supervised fine-tuned LLM,ensures that the generated text conforms to judicial writing conventions and formatting requirements.Experiments conducted on JuDGE,a benchmark for judicial document generation in the Chinese legal system,the results show that:1)MAC-AG consistently improves the document generation performance of seven baseline LLMs,with Qwen3-8B@MAC-AG achieving the most significant gains in sentencing prediction,charge determination,legal article citation,and semantic quality,demonstrating the generalizability of the proposed method;2)Compares with the state-of-the-art MRAG method,Qwen3-8B@MAC-AG improves the F1 score for charge prediction by 2.5%,the F1 score for legal article citation by 11.4%,and overall semantic similarity of judicial documents by 14.01%;3)Ablation experiments confirm that each agent is indispensable,as removing any component leads to a marked performance decline,highlighting the effectiveness of task decoupling and inter-agent collaboration;and 4)In human evaluation,Qwen3-8B@MAC-AG achieves an average score of 4.79/5.00,outperforming baseline methods across six dimensions:reasoning,logical consistency,norm compliance,completeness,readability,and value balancing.Overall,this study provides a reproducible multi-agent collaboration paradigm for judicial intelligence and substantially improves the accuracy of sentencing prediction,the rigor of legal reasoning,and the practical applicability of automated judicial document generation.

Semantics-aware Fine-grained Parallel Structural Reduction Framework for Petri-net LTL Model Checking

HE Yunong, DING Zhijun

Computer Science. 2026, 53 (8): 357-364. doi:10.11896/jsjkx.260100111

Abstract

PDF(2856KB) ( 84 )

References | Related Articles | Metrics

State space explosion severely limits the scalability of explicit-state LTL(Linear Temporal Logic) model checking for Petri nets.Structural reduction iteratively applies local structural transformations to remove redundant places and transitions,thereby reducing the size of subsequent product-automaton construction and counterexample-path search while preserving consistency in property decisions.To address the preprocessing bottleneck of traditional serial Scan-and-Commit loops on large-scale models,this paper proposes a semantics-aware,fine-grained parallel framework for structural reduction.By analyzing atomic-proposition references in LTL formulas,the framework constructs a PSS(Property Support Set) as the safety boundary for reduction,performs unified one-pass filtering during candidate generation based on the non-intersection between candidate scopes and the PSS,further uses an incremental Impact Set to drive local rescanning,and approximately solves the MIS(Maximum Indepen-dent Set) problem on the scope-conflict graph to schedule conflict-free reduction batches.Statistics over 410 public P/T-net instances from the MCC(Model Checking Contest) 2025 show that,under the K=24 configuration,highly structurally redundant model families achieve 2~4.5x reduction speedups,while most low-redundancy cases remain comparable to the serial baseline; performance regressions on a small number of short-running tasks are mainly attributable to fixed parallel overheads and the serialbarrier in the commit phase.For properties containing the Next(X) operator,the framework supports automatic fallback to a safe rule subset,strictly maintaining the boundary of semantic correctness.

Test Case Generation and Prioritization for Program Fault Diagnosis Based on Large LanguageModel

FAN Ruxin, SUN Baicai, GONG Lina, YAO Xiangjuan, GONG Dunwei

Computer Science. 2026, 53 (8): 365-374. doi:10.11896/jsjkx.250900162

Abstract

PDF(3098KB) ( 65 )

References | Related Articles | Metrics

In recent years,large language model(LLM) have been widely applied to test case generation and prioritization.However,existing studies show that LLM-generated test cases still suffer from high error rates and semantic deviations.Empirical evaluations indicate that only about 40% of the generated code is directly executable,with roughly one-third of the failures caused by assertion logic errors.These issues significantly undermine the effectiveness and reliability of LLM-based fault diagnosis.To address this problem,this paper proposes a fault diagnosis-oriented test case generation and prioritization method based on LLM.Firstly,the proposed approach extracts code summaries from the original program using an LLM to generate substitute programs that implement equivalent functionality.Then,the LLM is employed to produce initial test cases,whose expected outputs are refined through multiple substitute programs.Next,the differences between the outputs of the original and substitute programs are analyzed,and test cases are prioritized based on the function call relationships of the original program.Finally,program faults are diagnosed efficiently according to the prioritized sequence.Experiments are conducted on seven benchmark problems comprising 140 Python programs.The results demonstrate that the proposed method significantly reduces the error rate of LLM-generated test cases.The resulting test sequences achieve high fault coverage across multiple LLM,with an average fault detection rate(APFD) consistently above 0.75,confirming the proposed method’s effectiveness in improving test accuracy and fault diagnosis efficiency.

Review of Research on Face Deepfake Detection Technology

WANG Lihua, WANG Xinyu, YAN Weidan, ZHANG Dengyin

Computer Science. 2026, 53 (8): 375-387. doi:10.11896/jsjkx.251100118

Abstract

PDF(2497KB) ( 77 )

References | Related Articles | Metrics

With the rapid advancement of AI-generated content(AIGC),social media and other online platforms have been inundated with large volumes of high-quality face deepfake images and videos,which pose significant security threats to personal privacy protection and the social public opinion ecosystem.As a key means to address this issue,face deepfake detection technology has emerged as a research focus in recent years.This paper provides a comprehensive overview of existing technologies,extending from a data modality perspective to a learning paradigm perspective.Firstly,it systematically reviews the mainstream and representative image-level and video-level face forgery detection technologies.From the perspectives of feature dimensions and technical architectures,the former is categorized into methods based on the spatial domain,frequency domain,and spatial-frequency combination;the latter is categorized into methods based on biological features and spatio-temporal inconsistency.Secondly,since the collaborative relationships of multimodal information in real-world scenarios are difficult to be perfectly replicated,multimodal face deepfake detection technologies are further explored.Building on the analysis of audio forgery detection,the review focuses on methods based on cross-modal consistency and multi-modal feature fusion.Furthermore,the review shifts perspective to explore detection technologies oriented towards advanced learning paradigms.It highlights generalization and interpretability me-thods addressing cross-domain adaptability and decision transparency challenges,and analyzes the new paradigm of large model-aided detection.In addition,the mainstream face deepfake detection datasets and common evaluation metrics are summarized,thereby providing a reference for technical comparison and performance assessment.Finally,the limitations of the current face deepfake detection field are summarized,and prospects for future development directions are presented.

Survey on Code Representation Learning for Vulnerability Detection

LI Xiaochao, YUAN Zisu, LI Qianmu, LIU Fan, CHE Xun

Computer Science. 2026, 53 (8): 388-402. doi:10.11896/jsjkx.250900036

Abstract

PDF(1803KB) ( 66 )

References | Related Articles | Metrics

Software security is fundamental to the stable operation of the digital society,as cybersecurity incidents caused by software vulnerabilities result in substantial economic losses.Traditional detection methods struggle to manage the increasing scale and complexity of code.In recent years,deep learning-based vulnerability detection techniques have made remarkable progress.The core challenge lies in encoding the syntactic,semantic,and structural information of source code into low-dimensional,con-tinuous vector representations to enable effective processing and analysis by deep learning models—a process known as code representation learning.This paper systematically reviews code representation learning techniques for vulnerability detection.It categorizes techniques into sequence-level,graph-level,and fusion-level approaches,evaluates the performance of deep learning-based methods in vulnerability detection,and identifies key challenges like dataset imbalance,poor cross-domain generalization,and li-mited model interpretability.The study reveals that representation methods integrating sequential and graph structural features improve the average F1-score by 15%~26% compared to single-modality approaches.Moreover,combining prompt learning with graph-guided strategies significantly enhances the vulnerability detection capabilities of LLMs.This review provides a structured reference framework for the software security community and identifies promising future directions such as federated learning,explainable AI(XAI),and cross-language representation learning.

Reversible Anomaly Traffic Detection for Programmable Data Plane

SHU Yapeng, DU Yang, HUANG He, SUN Yu’e

Computer Science. 2026, 53 (8): 403-412. doi:10.11896/jsjkx.251100114

Abstract

PDF(3455KB) ( 76 )

References | Related Articles | Metrics

Meeting the demands of network management and security protection in high-speed network environments,real-time reversible traffic measurement has become a critical foundation for supporting tasks such as anomaly detection and attack tracing.Traditional reversible Sketch methods face dual challenges when deployed on P4-programmable data planes:the absence of multiplication,division,and modulo instructions in the P4 instruction set prevents the direct implementation of reversible Sketches based on the Chinese Remainder Theorem(CRT).Furthermore,the co-prime moduli required by CRT often conflict with hardware memory allocation granularity,leading to memory fragmentation,while the computational complexity of reverse recovery for multiple anomalous flows grows exponentially.To address these issues,an efficient abnormal traffic detection mechanism is designed.By constructing a special set of co-prime moduli,modulo operations are transformed into bit-shifts and additions/subtractions.A register pool with a complementary allocation strategy is established to eliminate memory fragmentation,supporting the runtime redeployment of measurement tasks.Additionally,a dual-Sketch coordination mechanism is introduced to reduce the computational complexity of reverse recovery from exponential to linear.Experimental results on a P4 switch demonstrate microsecond-level memory allocation(45~58 μs),millisecond-level task switching latency(approximately 80 ms for 8 tasks),and a processing delay of only 0.4 μs.It does not interfere with normal packet forwarding,which effectively validating its feasibility and efficiency in high-speed networks.

Semi-supervised Source Code Vulnerability Detection Model with Dual-stage Label Propagation

YAN Qisen, KANG Ming, YANG Yiyao, CHEN Wen

Computer Science. 2026, 53 (8): 413-425. doi:10.11896/jsjkx.250800088

Abstract

PDF(4786KB) ( 71 )

References | Related Articles | Metrics

Automated source code vulnerability detection is crucial for ensuring software security.Currently,deep learning-based vulnerability detection models predominantly rely on supervised learning,which requires a substantial number of credible,labeled vulnerability samples for training.However,the collection of such samples is costly and dependent on expert knowledge.More-over,real-world applications often face challenges such as class imbalance.To address these challenges,this paper proposes DLP-VulD,a novel semi-supervised model that introduces a dual-stage label propagation method to the field of source code vulnerability detection for the first time.Firstly,it extracts program slices containing vulnerability context from source code using code property graphs(CPG) and learns their semantic and structural features with a BiLSTM model to generate code feature vectors.Subsequently,DLP-VulD employs the dual-stage label propagation mechanism to extend the training set by propagating trusted label information from labeled to unlabeled samples.In the first stage of label propagation,leveraging an initial small set of labeled samples,DLP-VulD calculates label influence to select high-credible pseudo-labeled(HCP) samples and partial-labeled(PAL) samples(those with multiple candidate labels) from the unlabeled set.In the second stage,to mitigate label propagation noise,a multi-source partial label learning(MSPLL) method is proposed to perform label disambiguation for the PAL samples.Finally,in the vulnerability detection module,DLP-VulD utilizes the original labeled samples,the HCP samples,and the disambiguated PAL samples to form an augmented dataset for reinforced training,thereby enhancing the classifier’s detection performance.Additionally,DLP-VulD adopts a class-weighting mechanism to reduce the impact of class imbalance.Comparative experiments on multiple real-world open-source project datasets and synthetic vulnerability datasets demonstrate that DLP-VulD surpasses traditional supervised methods in overall performance,showing significant improvements in Recall and F1-score compared to baselines such as Devign,VulDeePecker,and LineVul.On six real-world vulnerability datasets characterized by class imbalance and diverse vulnerability types,DLP-VulD achieves an average reduction of 8% in false positive rate(FPR) and 29.6% in false negative rate(FNR) compared to the baselines.Even with a very limited labeled set(10%),DLP-VulD effectively balances its detection capability for true vulnerabilities with control over false positives,achieving superior comprehensive performance.Experimental results confirm the effectiveness of DLP-VulD for training and deploying vulnerability detection models in environments with scarce labeled data.

Network Anomaly Traffic Detection Based on Deep Multi-instance Learning

FENG Haoyu, ZHANG Yuxuan, LIU Zixuan, MENG Hua

Computer Science. 2026, 53 (8): 426-436. doi:10.11896/jsjkx.250600041

Abstract

PDF(5461KB) ( 75 )

References | Related Articles | Metrics

With the exponential growth of network scale and the increasing complexity of traffic patterns,traditional anomaly traffic detection methods based on individual traffic samples face serious challenges in terms of real-time performance and computational efficiency.To address this issue,this study proposes a deep multi-instance learning framework with a gated attention mechanism,named GAD-MIL.This framework processes traffic samples in bags,replacing traditional instance-level analysis with bag-level analysis and anomaly localization to achieve efficient abnormal traffic detection.Specifically,for the bagged traffic data,the proposed method adopts a two-stage learning architecture.Firstly,a pre-trained feature extractor is used to generate discriminative traffic embeddings.Secondly,a gated attention-based multi-instance pooling layer is introduced to dynamically aggregate instance features within each bag and identify anomalous samples.This architecture overcomes traditional MIL models’ reliance on bag-level labels,enabling end-to-end instance-level anomaly localization while maintaining low computational complexity.Experimental results on five benchmark datasets such as CICIDS2017 and DoH2020 show that,GAD-MIL significantly reduces in-ference time and achieves a 2.72 percentage-point improvement in F1 score over traditional deep learning methods on the CICIDS2017 dataset and a 4.2× speedup in computational efficiency.

Research on Deep Learning-based Side-channel Analysis Method with Dynamically ComposableMulti-head Attention

JIANG Lingla, CHEN Wen, SUN Wei, ZHAO Kui

Computer Science. 2026, 53 (8): 437-445. doi:10.11896/jsjkx.250600213

Abstract

PDF(2651KB) ( 66 )

References | Related Articles | Metrics

Side-channel analysis faces the challenge of extracting key-related features from a large number of power traces.Since MHA(Multi-Head Attention) mechanism enables multiple heads to capture both local dependencies and global correlations in data,it has strong multi-feature learning capabilities.Therefore,MHA has been widely applied to automatic feature extraction in side-channel analysis in recent years.However,when multiple heads learn features simultaneously,MHA is prone to the limitations of low-rank bottlenecks and redundant heads.This problem weakens the ability of MHA to capture complex temporal dependencies and global feature relationships in data,leading to instability during training and difficulty in converging to the optimal solution.To address this issue,this paper proposes a deep side-channel analysis method based on dynamically composable multi-head attention.The method introduces a dynamically composable multi-head attention mechanism that adaptively combines information from different attention heads,thereby effectively enhancing the model’s capability to extract key features,ensuring training stability,and continuously improving attack performance.Comparative experiments conducted on the public ASCAD,AES_HD,and CHES18 datasets demonstrate that the proposed method outperforms existing models in both training stability and attack effectiveness.For example,on the AES_HD and CHES18 datasets,the number of power traces required for successful attacks is reduced by 56.2% and 66.7%,respectively.

Program Capabilities Reduction Based on Code Specialization

FAN Yuhao, SUN Cong, ZHANG Linmao

Computer Science. 2026, 53 (8): 446-454. doi:10.11896/jsjkx.250600180

Abstract

PDF(2285KB) ( 71 )

References | Related Articles | Metrics

The setuid mechanism in Linux enables users to temporarily escalate the privileges of programs.Such a mechanism violates the principle of least privilege.Once vulnerabilities exist in these programs,attackers can exploit the privileges to conduct privilege escalation,resulting in severe consequences.To address this issue,the capability mechanism in Linux decomposes the root privilege into a set of fine-grained capabilities and assigns only the required capability to the program,thereby mitigating risks.Existing capability-limiting approaches based on system-call identification have exhibited drawbacks when obtaining the genuine capability subset required by the program.On one hand,due to the failure to filter out redundant code in dynamic libraries and the inaccurate analysis of the Glibc call graph,the resulting system-call set is over-approximated.On the other hand,the inaccurate system-call parameter analysis leads to a conservative estimation of capabilities for the system calls.This paper presents a program capability-limiting framework based on code specialization.The program-dependent dynamic libraries are pruned to reduce the impact of redundant library code on the system-call identification.Combined with the dynamic library pruning results,this work analyzes the system calls on the pruned Glibc call graph and obtains a more precise set of system calls specific to the program.Moreover,this work conducts iterative backward data-flow analysis on the pruned program-specific dynamic libraries to extract sensitive parameter values for specific system calls.Based on the program-specific system call set and the sensitive para-meter values,the accurate capability set required by the program is identified.For a specific binary program,the proposed framework conducts the analysis and restricts the program with the accurate capability set and system-call set,thereby reducing the privileged operations that the program can perform.Experimental results demonstrate that,compared to state-of-the-art approaches,the proposed approach identifies system calls more accurately and enforces stronger capability restrictions on the program.On setuid programs,this work can significantly reduce the attack surface and alleviate privilege-escalation vulnerabilities.

Multi-aggregation Heterogeneous Graph Neural Network-based Saturation Attack DetectionMethod for SDN

QIAN Qing, RAN Longwen, WANG Huan, CUI Yunhe, WANG Lingyun

Computer Science. 2026, 53 (8): 455-468. doi:10.11896/jsjkx.251000056

Abstract

PDF(3498KB) ( 72 )

References | Related Articles | Metrics

Saturation attack is a resource-consuming attack and one of the key cybersecurity threats facing software-defined networks.The current saturation attack detection methods based on graph neural networks mainly model network flows as isomorphic structures and achieve certain results,However,they fail to fully capture the complex interaction characteristics between network devices and traffic during the detection process.At the same time,the existing methods have limitations in the node information aggregation mechanism,which limits the further improvement of their detection accuracy.To solve the above problems,this paper proposes a saturation attack detection method based on a heterogeneous hierarchical graph neural network,MARATHON.This method constructs a heterogeneous flow graph structure that can accurately characterize the relationship between traffic,host and switch,which fully considers the differentiated characteristics of traffic,host and switch,so as to effectively describe the interaction behavior between devices and traffic.On this basis,a new graph neural network detection model is designed,which integrates three types of aggregation mechanisms:host/switch isomorphic aggregation,host-traffic/switch-traffic heterogeneous aggregation,and traffic-host-traffic/traffic-switch-traffic metapath aggregation,which fully mines the potential discriminant information in the network flow through hierarchical feature extraction strategy to improve the accuracy of attack detection.Experimental results on the public datasets IMC10 and 4SICS show that the MARATHON method is better than the existing mainstream methods in terms of accuracy,recall and F1 score.Specifically,under the Fattree topology of IMC10 dataset,the accuracy and macro recall of MARATHON reach 97.5% and 98.3%,respectively.Under the Agis topology of the 4SICS dataset,the accuracy and macro recall are further improved to 98.4% and 98.5%,which verifies the effectiveness and superiority of the proposed method.

Dynamic Adversarial Detection Framework Based on Multimodal Uncertainty Fusion

FU Le, HUANG Xiaofang, LIAO Min, SONG Luhua

Computer Science. 2026, 53 (8): 469-477. doi:10.11896/jsjkx.250700031

Abstract

PDF(2345KB) ( 73 )

References | Related Articles | Metrics

Aiming at the problem that the intrusion detection system based on deep neural network shows insufficient robustness under diverse adversarial attacks and the traditional fusion method is difficult to take into account the characteristics of different attack scenarios,this paper proposes a dynamic adversarial sample detection framework based on multi-modal uncertainty fusion.Firstly,a training progress aware dynamic loss scheduling mechanism is designed to realize the balance optimization between feature learning and defense robustness by linearly scheduling the weights of normal samples and adversarial samples.Secondly,a collaborative fusion mechanism of uncertainty and attention is proposed.Monte Carlo dropout is used to construct the prediction distribution and quantify the uncertainty of the model.Then,the context-sensitive attention module is combined to adaptively adjust the contribution weights of each sub-model to realize the dynamic weighted integration of multimodal features,so as to effectively suppress the anti-noise and highlight the key features.Experimental results show that the proposed framework achieves more than 96% detection accuracy and nearly 100% AUC under normal samples and multiple attack samples(FGSM,PGD,BIM),which is significantly better than AT-DNN and Roshan and other representative adversarial defense methods,fully veri-fying the effectiveness of the proposed framework in improving robustness and generalization ability.