Table of Content

30 January 2026, Volume 44 Issue 1

Previous Issue   

Special Issue on Computer Application

Network Community Detection Based on Structure-Enhanced Deep Clustering

LI Yongzhen, MA Fuyuan, MA Shixuan, WANG Yuhan, WANG Ying

2026, 44(1):  1-20.  doi:10.3969/j.issn.0255-8297.2026.01.001

Asbtract ( 54 )   PDF (3512KB) ( 26 )  

References | Related Articles | Metrics

Community detection in social networks is important for applications such as information diffusion, recommendation, and advertising. However, existing methods still face challenges in feature fusion, sparse graph modeling, and multi-source information utilization. To address these issues, this paper proposed a structure-enhanced deep clustering (SDC) model for network community detection, which consisted of four key modules. First, the topology enhancement module built an enhanced adjacency matrix by modeling second-order similarity between nodes, which alleviated the problem of missing high-order relations in sparse social networks. Second, the multi-view feature fusion module dynamically fused node attributes and topology features at the node level, and integrated semantic information from both the original and enhanced graphs at the graph level. Third, the multi-source distribution fusion clustering module used learnable weights to integrate clustering information from different feature spaces at the distribution level, balancing local topology and global semantics. Finally, the dual self-supervised module optimized the model through Kullback-Leibler (KL) divergence alignment, node reconstruction, and similarity constraints. Experiments show that compared with mainstream baseline methods, SDC model improves ACC, NMI, ARI, and F1 by an average of 3.80%, 9.09%, 11.21%, and 7.43%, respectively on the three benchmark datasets. Simulations based on Facebook interaction data also demonstrate the ability of SDC model to capture community structure evolution.

Real-Time Travel Pattern Recognition Algorithm Based on Self Adaptive Pooling Enhanced Attention Mechanism

LI Yinxiang, DU Wenyuan, XU Zhe, PENG Chen, YAN Jianqiang

2026, 44(1):  21-33.  doi:10.3969/j.issn.0255-8297.2026.01.002

Asbtract ( 47 )   PDF (1311KB) ( 23 )  

References | Related Articles | Metrics

Identifying travel patterns is a crucial task in integrated mobility as a service. To address the limitations of existing traffic travel pattern recognition algorithms, namely insufficient accuracy and real-time application demands, this paper proposed a real-time travel pattern recognition algorithm based on a self adaptive pooling enhanced attention mechanism. Based on the point cloud network, the proposed algorithm efficiently learned current and historical information by incorporating causal convolution and causal pooling, thereby achieving real-time travel pattern recognition. A self adaptive pooling enhanced attention module was further embedded into the framework model to calculate the weight map among different features, thereby enhancing the feature modeling capability. Additionally, the algorithm integrated both motion and geographical features, which effectively improved the recognition accuracy of bus and car travel patterns. Experimental results show that the proposed algorithm achieves superior accuracy. Compared with other one-stage methods, its recognition accuracy is improved by approximately 0.05. Compared with the latest two-stage models such as FPbiLSTM, the parameter count of the proposed algorithm is only 0.167 that of these models, making the proposed approach more lightweight and suitable for deployment on mobile devices.

Complex Logical Query Model Based on Improved Transformer

CHEN Yuyin, LI Guanfeng, QIN Jing, XIAO Yuhang

2026, 44(1):  34-49.  doi:10.3969/j.issn.0255-8297.2026.01.003

Asbtract ( 40 )   PDF (819KB) ( 11 )  

References | Related Articles | Metrics

With the widespread application of knowledge graphs in scenarios such as intelligent question answering and recommendation systems, answering complex logical queries on incomplete knowledge graphs has become the focus and difficulty of current research. In view of the fact that ordinary embedding-based methods need to be trained on complex logical queries and cannot be well generalized to query structures outside the distribution, this paper proposed an improved-Transformer-based model DCMHA-MoE that integrated the dynamically composable multi-head attention (DCMHA) mechanism and the mixture-of-experts (MoE) network. This model represented complex query graphs as sequence inputs through triple transformation and bidirectional path encoding technology, and dynamically modeled the structural dependencies and semantic interactions therein, so that complex logical queries can be realized. The DCMHA realized the adaptive combination of attention heads to enhance the semantic expression ability. The MoE network introduced a sparse activation mechanism to improve the adaptability to different query structures and reduce the computational cost. Experiments were conducted on the FB15K-237 and NELL-995 datasets. The results show that compared with the baseline model DiffCLR, the DCMHA-MoE model improves the mean reciprocal rank (MRR) in existential positive first-order logic (EPFO) query $(\wedge, \vee)$ by 10.4% and 7.2%, respectively, which verifies the effectiveness and superiority of DCMHA-MoE in complex logical query tasks.

Uncertain Knowledge Graph Reasoning Model Based on Gaussian Metric Learning

ZHANG Yuting, TENG Fei, YE Xiaoqing

2026, 44(1):  50-66.  doi:10.3969/j.issn.0255-8297.2026.01.004

Asbtract ( 39 )   PDF (977KB) ( 14 )  

References | Related Articles | Metrics

Long-tail relations containing only a small number of facts are prevalent in real-world knowledge graphs, and few-shot knowledge graph completion aims to address this data sparsity issue. However, existing approaches often neglect the inherent uncertainty of entities and triples, which limits model reasoning performance under noisy or data-scarce conditions. This paper proposed a covariance-optimized Gaussian metric learning model for uncertain completion (CoGMUC) model to tackle uncertain reasoning in few-shot knowledge graphs. This model represented entities and relations within knowledge graphs as Gaussian distributions, effectively capturing their inherent uncertainty through covariance matrices. It computed semantic similarity with a covariance-aware multi-matching network to complete missing facts and predict confidence levels. Furthermore, a difficult negative sample mining strategy was introduced to enhance the discriminative capability and generalization performance of the model. Experimental results on the public datasets NL27K and CN15K demonstrate that compared with the existing few-shot uncertain knowledge graph completion model based on Gaussian metric learning, CoGMUC improves mean reciprocal rank (MRR) by 21.8% and 2.3% and increases Hits@10 by 9.6% and 21.5%, respectively in the link prediction task. Meanwhile, in the confidence prediction task, the mean squared error (MSE) is reduced by 14.3% and 7.7%, respectively. The findings demonstrate that the CoGMUC model effectively models and leverages uncertainty information, significantly enhancing the performance of few-shot knowledge graph completion.

Construction of Malware Knowledge Graph for Threat Intelligence Analysis

XIANG Ga, HU Yan, ZHANG Yangsen, SUN Lu, QI Rui, TAN Zicheng

2026, 44(1):  67-82.  doi:10.3969/j.issn.0255-8297.2026.01.005

Asbtract ( 39 )   PDF (852KB) ( 5 )  

References | Related Articles | Metrics

Threat intelligence analysis is a crucial means to enhance proactive defense capabilities. Research on the construction of malware knowledge graphs holds significant importance for improving malware detection capabilities. In the construction of malware knowledge graphs, the accuracy and completeness of entity and relation extraction still require further improvement. This paper proposed a method for constructing malware knowledge graphs based on a joint extraction model. Firstly, a malware ontology model was proposed for threat intelligence analysis, defining 12 types of relations to standardize the expression of key knowledge about malware. Then, a joint extraction model based on RoBERTa with whole word masking (RoBERTa-Wwm) and pointer annotation was proposed to extract malware entities and their relations, thereby constructing a graph. The experiment demonstrates that the model achieves good performance with an F1 value of up to 0.841. This study is of great significance for the automatic analysis of malware threat intelligence, laying the foundation for improving proactive defense capabilities.

Federated Recommendation Algorithm Integrating Graph Neural Networks and Depth Graph Clustering

YI Huawei, SONG Shixi, WANG Yanfei, BAI Siyi

2026, 44(1):  83-96.  doi:10.3969/j.issn.0255-8297.2026.01.006

Asbtract ( 39 )   PDF (1065KB) ( 10 )  

References | Related Articles | Metrics

Federated learning, as the main framework for addressing privacy and security issues in recommendation systems, faces problems such as poor recommendation accuracy, insufficient privacy protection, and excessive communication overhead in practical applications. To address these issues, this paper proposed a federated recommendation algorithm integrating graph neural networks and depth graph clustering. Firstly, a graph neural network was used to capture high-order complex user and item interaction relationships, improving the recommendation accuracy of the recommendation system. Secondly, differential privacy noise was injected into the communication link between the federated learning clients and server to blur the true gradient, thereby enhancing the privacy protection capability of the recommendation system. Finally, clients were clustered by introducing depth graph clustering, and client representatives from each cluster were selected to participate in training. The obtained parameters were shared within the cluster to accelerate the convergence speed of the model and reduce communication overhead under the federated learning framework. The experimental results on real datasets show that the proposed algorithm can enhance privacy protection of the system and reduce communication overhead while improving recommendation accuracy.

Forest Image Dehazing Based on Feature Fusion Attention and Contrastive Learning

WU Wenqiang, CHEN Aibin, LI Xiaoyao

2026, 44(1):  97-109.  doi:10.3969/j.issn.0255-8297.2026.01.007

Asbtract ( 36 )   PDF (8033KB) ( 13 )  

References | Related Articles | Metrics

Current algorithms often struggle with incomplete dehazing, loss of details in dark areas, and color distortion in forest foggy images. To address these issues, this paper proposed an adaptive forest image dehazing algorithm based on feature fusion attention and contrastive learning. A multi-scale feature fusion attention mechanism was designed, which dynamically adjusted feature responses by combining channel and spatial attention, thereby enhancing the representation capability of important features. A local contrast regularization module was constructed to enhance the ability of the model to discriminate variations in fog concentration in dark and distant areas. Furthermore, an adaptive color correction module was introduced to mitigate color distortion. Experimental results on both synthetic and real-world forest foggy image datasets demonstrate that the proposed algorithm outperforms existing methods, achieving significant improvements in peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) and a reduction in natural image quality evaluator (NIQE), and exhibiting strong robustness and generalization ability.

Explainable Learning Path Recommendation Based on Dynamic Attention Reinforcement Learning

ZHANG Xiaoming, FENG Zejia, WANG Huiyong, ZHANG Xiaojing

2026, 44(1):  110-133.  doi:10.3969/j.issn.0255-8297.2026.01.008

Asbtract ( 44 )   PDF (1455KB) ( 13 )  

References | Related Articles | Metrics

The popularization of large-scale online education has made it difficult for learners to choose courses, and personalized learning path recommendation faces the challenge of relying on single modal data, which leads to the limitation of semantic representation. Moreover, static knowledge maps are difficult to generate dynamic explainable recommendation logic. To address the aforementioned issues, this paper proposed a framework of explainable learning path recommendation based on dynamic attention reinforcement learning (ELPR-DARL). Firstly, a heterogeneous collaborative knowledge graph was constructed, integrating course text, visual content, and knowledge dependencies to enhance cross-modal semantic alignment capabilities. Secondly, a dynamic attention aggregation mechanism for adjacent nodes was designed, which adjusts the weights of entity relationships through a bias correction strategy, and a bidirectional interaction aggregator was utilized to fuse multi-level neighborhood features, enhancing the fine-grained expression ability of knowledge reasoning. Finally, a knowledge graph-aware reinforcement learning strategy was proposed, which explicitly modelled the association between user behavior and knowledge topology based on path connectivity reward functions, generating explainable paths that include global rewards and local attention weights. Experiments based on the MOOC dataset show that this method achieves 22.85%, 33.81%, 52.01%, and 6.34% in NDCG, Recall, HR, and precision metrics, respectively, which is 2.88%, 3.55%, 2.42%, and 3.26% higher than the suboptimal model. User research shows that 80.36% of learners believe that path explanation significantly improves recommendation transparency. This study verifies that the collaborative optimization of a dynamic attention mechanism and reinforcement learning can effectively balance recommendation accuracy and explainability.

Three-Dimensional Fuzzy Clustering Algorithm Integrating Spatial Texture Features

JIN Zhengyang, YAN Shaohong, ZHANG Yanbo, YAO Xulong, TAO Zhigang, CHEN Zhiyuan

2026, 44(1):  134-148.  doi:10.3969/j.issn.0255-8297.2026.01.009

Asbtract ( 38 )   PDF (7112KB) ( 9 )  

References | Related Articles | Metrics

The traditional fuzzy C-means (FCM) clustering algorithm is highly sensitive to the initial cluster centers and the noise points. These limitations become more pronounced in complex environments or high-dimensional spaces. To overcome these issues, this study proposed a three-dimensional FCM algorithm integrating spatial texture features. The algorithm was designed to identify regions with noticeable density differences caused by uneven distribution of internal components in the analyzed objects. First, the method extended the two-dimensional gray-level co-occurrence matrix and planar texture feature theory into three-dimensional space to describe spatial texture features. Next, contrast texture features were used to improve the selection of initial cluster centers. Finally, dissimilarity texture features were integrated into the conventional objective function of FCM algorithm to enhance noise resistance. In a simulated experiment on fracture extraction, the proposed algorithm achieved an accuracy of 99.39%, representing a 34% improvement over the traditional FCM algorithm (accuracy of 65.31%). These results confirm the effectiveness of the new algorithm in extracting regions with noticeable density differences inside the analyzed objects. In practical applications, the new algorithm shows superior performance in identifying and extracting human thoracic skeleton.

Multi-granularity Semantic Aspect-Based Sentiment Analysis Model with Fusion of BERT Encoding Layers

XU Kai, CHI Mingde, WANG Qi, LI Jianzhou, ZHANG Hui

2026, 44(1):  149-165.  doi:10.3969/j.issn.0255-8297.2026.01.010

Asbtract ( 45 )   PDF (1725KB) ( 18 )  

References | Related Articles | Metrics

Aspect-based sentiment analysis (ABSA) aims to identify the sentiment polarity toward specific aspects within a text. However, existing research still faces multiple challenges: BERT-based approaches suffer from semantic overfitting and insufficient utilization of low-level semantic features; the self-attention mechanism is prone to losing local information; structures with multiple encoding layers and multi-granularity semantics lead to information redundancy. To address these issues, this paper proposed a multi-granularity semantic aspect-based sentiment analysis model with fusion of BERT encoding layers (MSBEL). The model introduced a pyramid attention mechanism to leverage semantic features from various encoding layers, and was combined with low-level encoders to mitigate overfitting. It employed multi-scale gated convolution to enhance its capability in handling local information loss and utilized cosine attention to highlight sentiment features relevant to aspect terms, thereby reducing information redundancy. t-SNE visualization demonstrates that the clustering effect of sentiment representations of MSBEL is superior to that of BERT. MSBEL was compared with mainstream models on multiple benchmark datasets. Compared with LCF-BERT, it achieves F1 improvements of 1.53%, 3.94%, 1.39%, 6.68%, and 5.97% on five datasets. In comparison with SenticGCN, it achieves an average increase of F1 by 0.94% and a maximum increase of 2.12%. Compared with ABSA-DeBERTa, MSBEL increases the F1 by 1.16% on average and achieves a maximum increase of 4.20%. These results validate the effectiveness and superiority of the proposed model for ABSA tasks.

Completion Method for Ship Point Cloud Based on Symmetry Priors

ZENG Yinchuan, ZHENG Bo, WANG Xianbao, XIANG Sheng

2026, 44(1):  166-180.  doi:10.3969/j.issn.0255-8297.2026.01.011

Asbtract ( 44 )   PDF (2289KB) ( 10 )  

References | Related Articles | Metrics

Due to the inherent limitations of single-view scanning and the spatial occlusion effects of complex ship hull structures, existing data collection systems commonly face the technical bottleneck of extensive missing data in back-side point clouds. To address this challenge, this paper proposed a ship point cloud completion method based on symmetry priors. This method operated without the need for labeling data and utilized the symmetrical structural characteristics of ships as prior-driven knowledge to effectively complete the back-side point clouds. First, a feature extraction model of the longitudinal centerplane for various types of ship hull was established based on geometric topology analysis of the ships. Then, a symmetry transformation field generation algorithm was proposed to make a mirror completion for the missing point clouds along the longitudinal centerplane of the ship hull, thereby constructing a candidate point cloud set for completion. Finally, an average nearest neighbor quality assessment function between the candidate point clouds and the original point clouds was designed to robustly select the optimal completion result. Experimental results show that the proposed method effectively completes the back-side point clouds of typical ship types, such as sharp-prowed and flat-bottomed ships, without requiring any training samples, and it meets the requirements of real-time data collection scenarios.