Knowledge-Based Systems最新文献_第10页

Cross-scale channel attention and feature fusion-aware aggregation for sonar images object detection 声纳图像目标检测中的跨尺度通道关注和特征融合感知聚合

IF 7.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Knowledge-Based Systems

Pub Date : 2026-01-20 DOI: 10.1016/j.knosys.2026.115371

Pengfei Shi , Hanren Wang , Qianqian Zhang , Yuanxue Xin

Feature extraction and feature fusion are crucial for sonar image target detection. In terms of feature extraction, due to device limitations and the complexity of the underwater environment, sonar images often exhibit high levels of noise, which results in high similarity between targets and background, thus affecting feature extraction. In terms of feature fusion, transformer-based models rely on self-attention mechanisms, but this leads to a lack of local prior information. The interference from noise and the similarity between targets and background disrupt the computation of global relationships, confusing noisy features with useful ones, leading to insufficient geometric information and ultimately affecting detection accuracy. To address these issues, we propose an advanced detection framework that combines effective feature extraction and multi-scale feature fusion. We introduce a cross-scale channel attention module that dynamically adjusts channel weights by integrating the advantages of the squeeze-and-excitation (SE) module and the efficient multi-scale attention (EMA) module, capturing multi-scale dependencies, suppressing background noise, and enhancing global feature representation. Moreover, to further improve the effectiveness of feature fusion and better leverage geometric information, we design a CNN-based feature fusion perception aggregation network. This network promotes interaction between low-level geometric details and high-level semantic information through skip connections, enhancing feature representation and improving detection accuracy. Experimental results show that our method outperforms some advanced detection models in terms of detection performance.

特征提取和融合是声纳图像目标检测的关键。在特征提取方面，由于设备的限制和水下环境的复杂性，声纳图像往往表现出高水平的噪声，导致目标与背景高度相似，从而影响特征提取。在特征融合方面，基于变压器的模型依赖于自关注机制，但这导致缺乏局部先验信息。噪声的干扰以及目标与背景的相似性破坏了全局关系的计算，使噪声特征与有用特征混淆，导致几何信息不足，最终影响检测精度。为了解决这些问题，我们提出了一种结合有效特征提取和多尺度特征融合的高级检测框架。我们引入了一个跨尺度通道注意模块，该模块通过集成压缩激励（SE）模块和高效多尺度注意（EMA）模块的优点，动态调整通道权重，捕获多尺度依赖关系，抑制背景噪声，增强全局特征表示。此外，为了进一步提高特征融合的有效性，更好地利用几何信息，我们设计了一个基于cnn的特征融合感知聚合网络。该网络通过跳过连接促进低级几何细节与高级语义信息的交互，增强特征表示，提高检测精度。实验结果表明，该方法在检测性能上优于一些先进的检测模型。

{"title":"Cross-scale channel attention and feature fusion-aware aggregation for sonar images object detection","authors":"Pengfei Shi , Hanren Wang , Qianqian Zhang , Yuanxue Xin","doi":"10.1016/j.knosys.2026.115371","DOIUrl":"10.1016/j.knosys.2026.115371","url":null,"abstract":"<div><div>Feature extraction and feature fusion are crucial for sonar image target detection. In terms of feature extraction, due to device limitations and the complexity of the underwater environment, sonar images often exhibit high levels of noise, which results in high similarity between targets and background, thus affecting feature extraction. In terms of feature fusion, transformer-based models rely on self-attention mechanisms, but this leads to a lack of local prior information. The interference from noise and the similarity between targets and background disrupt the computation of global relationships, confusing noisy features with useful ones, leading to insufficient geometric information and ultimately affecting detection accuracy. To address these issues, we propose an advanced detection framework that combines effective feature extraction and multi-scale feature fusion. We introduce a cross-scale channel attention module that dynamically adjusts channel weights by integrating the advantages of the squeeze-and-excitation (SE) module and the efficient multi-scale attention (EMA) module, capturing multi-scale dependencies, suppressing background noise, and enhancing global feature representation. Moreover, to further improve the effectiveness of feature fusion and better leverage geometric information, we design a CNN-based feature fusion perception aggregation network. This network promotes interaction between low-level geometric details and high-level semantic information through skip connections, enhancing feature representation and improving detection accuracy. Experimental results show that our method outperforms some advanced detection models in terms of detection performance.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"337 ","pages":"Article 115371"},"PeriodicalIF":7.6,"publicationDate":"2026-01-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080658","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Dual prompts guided cross-domain transformer for unified day-night image dehazing 双提示引导跨域变压器统一昼夜图像去雾

IF 7.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Knowledge-Based Systems

Pub Date : 2026-01-20 DOI: 10.1016/j.knosys.2026.115362

Jianlei Liu , Jiaming Niu , Xiang Chen , Yuting Pang , Shilong Wang

Although considerable progress has been made in image dehazing, most existing methods are constrained to a single degradation type or specific haze pattern. However, in real-world environments, haze manifests in diverse forms owing to variations in illumination, day-night transitions, and other coupled degradation factors. A new task has been assigned to address the following challenges: unified day-night image dehazing (UDND), with the aim to restore haze-degraded images across daytime and nighttime conditions within a single unified framework. For this task, we propose UDNDformer, a cross-domain Transformer guided by dual-prompt learning, which integrates both hard prompt learning (HPL) and soft prompt learning (SPL). The HPL module reconstructs scene before encoding transferable haze representations in a frozen form, ensuring consistent degradation modeling across domains. By contrast, the SPL module employs learnable tensors that interact with encoded features to adaptively capture temporal haze variations and dynamically modulate restoration during decoding for condition-aware guidance. This dual-prompt design enables UDNDformer to achieve adaptive haze perception and flexible degradation modeling under diverse illumination conditions, thereby markedly enhancing the restoration quality in unified day-night scenarios. Extensive experimentation demonstrates that UDNDformer consistently outperforms state-of-the-art methods across multiple day-night benchmarks and demonstrates notable improvements in downstream vision tasks, validating its effectiveness and strong generalizability to real-world applications.

尽管在图像去雾方面已经取得了相当大的进展，但大多数现有的方法都局限于单一的降解类型或特定的雾霾模式。然而，在现实环境中，由于光照变化、昼夜转换以及其他耦合降解因素，雾霾表现为多种形式。一项新的任务被分配来应对以下挑战：统一的昼夜图像去雾（UDND），目的是在一个统一的框架内恢复白天和夜间条件下的雾退化图像。为此，我们提出了一种基于双提示学习的跨域转换器UDNDformer，它集成了硬提示学习（HPL）和软提示学习（SPL）。HPL模块在以冻结形式编码可转移的雾霾表示之前重建场景，确保跨域一致的退化建模。相比之下，SPL模块采用可学习的张量，与编码特征相互作用，自适应捕获时间雾霾变化，并在解码过程中动态调制恢复，以实现条件感知制导。这种双提示设计使UDNDformer能够在不同光照条件下实现自适应雾霾感知和灵活的退化建模，从而显著提高统一日夜场景下的恢复质量。大量的实验表明，UDNDformer在多个昼夜基准测试中始终优于最先进的方法，并在下游视觉任务中表现出显著的改进，验证了其有效性和对现实世界应用的强大通用性。

{"title":"Dual prompts guided cross-domain transformer for unified day-night image dehazing","authors":"Jianlei Liu , Jiaming Niu , Xiang Chen , Yuting Pang , Shilong Wang","doi":"10.1016/j.knosys.2026.115362","DOIUrl":"10.1016/j.knosys.2026.115362","url":null,"abstract":"<div><div>Although considerable progress has been made in image dehazing, most existing methods are constrained to a single degradation type or specific haze pattern. However, in real-world environments, haze manifests in diverse forms owing to variations in illumination, day-night transitions, and other coupled degradation factors. A new task has been assigned to address the following challenges: unified day-night image dehazing (UDND), with the aim to restore haze-degraded images across daytime and nighttime conditions within a single unified framework. For this task, we propose UDNDformer, a cross-domain Transformer guided by dual-prompt learning, which integrates both hard prompt learning (HPL) and soft prompt learning (SPL). The HPL module reconstructs scene before encoding transferable haze representations in a frozen form, ensuring consistent degradation modeling across domains. By contrast, the SPL module employs learnable tensors that interact with encoded features to adaptively capture temporal haze variations and dynamically modulate restoration during decoding for condition-aware guidance. This dual-prompt design enables UDNDformer to achieve adaptive haze perception and flexible degradation modeling under diverse illumination conditions, thereby markedly enhancing the restoration quality in unified day-night scenarios. Extensive experimentation demonstrates that UDNDformer consistently outperforms state-of-the-art methods across multiple day-night benchmarks and demonstrates notable improvements in downstream vision tasks, validating its effectiveness and strong generalizability to real-world applications.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"337 ","pages":"Article 115362"},"PeriodicalIF":7.6,"publicationDate":"2026-01-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146015908","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Fact splices and entity aggregation networks for sparse temporal knowledge graph completion 稀疏时态知识图补全的事实拼接和实体聚合网络

IF 7.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Knowledge-Based Systems

Pub Date : 2026-01-20 DOI: 10.1016/j.knosys.2026.115387

Lin Zhu , Jiahui Hu , Luyi Bai

Advancements in artificial intelligence has markedly highlighted the importance of temporal knowledge graphs. However, since factors such as limitations in data collection and the immaturity of knowledge extraction techniques, temporal knowledge graphs often remain incomplete. Moreover, the rapid growth of real-world information has significantly exacerbated the sparsity of these graphs, severely affecting their practical application effects. Precisely for these reasons, the enhancement of sparse temporal knowledge graphs has become a significant focus of investigation in current academic research. In the context of completing sparse temporal knowledge graphs, previous methods have enriched entity representations through neighbor information aggregation, alleviated the sparsity of the graphs, and improved the completion effect. However, these methods have limitations. On the one hand, they focus on the information of neighboring entities and neglect the role of the relationship vectors between the current entity and its neighbor entities. On the other hand, they fail to distinguish the aggregation weights according to the roles of adjacent entities, thus limiting the further improvement of the completion effect. Accordingly, this paper investigates an information aggregation method based on relevant facts and a role-oriented attention network to enrich entity representations. Given that the importance of relationship vectors is often overlooked, we propose a fact vector generation strategy through a chained relation extractor and a fact vector splicer to excavate the information of relationship vectors. Aiming at the problem that previous methods failed to distinguish the role weights of adjacent entities, we propose a role-oriented attention network. This network aggregates context information and assigns weights according to the roles of the aggregated information, thereby generating more accurate entity representations. According to the experimental results, our model outperforms state-of-the-art baseline models in the selected metrics.

人工智能的进步显著地突出了时间知识图的重要性。然而，由于数据收集的局限性和知识提取技术的不成熟等因素，时间知识图往往是不完整的。而且，现实世界信息的快速增长显著加剧了这些图的稀疏性，严重影响了它们的实际应用效果。正是由于这些原因，对稀疏时态知识图的增强已经成为当前学术研究的一个重要研究热点。在完成稀疏时态知识图的背景下，以前的方法通过邻居信息聚合丰富了实体表示，减轻了图的稀疏性，提高了完成效果。然而，这些方法有局限性。一方面，它们关注的是相邻实体的信息，而忽略了当前实体与其相邻实体之间的关系向量的作用。另一方面，未能根据相邻实体的角色区分聚合权值，限制了补全效果的进一步提高。为此，本文研究了一种基于相关事实的信息聚合方法和一个面向角色的注意网络来丰富实体表征。针对关系向量的重要性经常被忽视的问题，提出了一种通过链式关系提取器和事实向量拼接器来挖掘关系向量信息的事实向量生成策略。针对以往方法无法区分相邻实体的角色权重的问题，提出了一种面向角色的关注网络。该网络聚合上下文信息，并根据聚合信息的角色分配权重，从而生成更准确的实体表示。根据实验结果，我们的模型在选定的指标中优于最先进的基线模型。

{"title":"Fact splices and entity aggregation networks for sparse temporal knowledge graph completion","authors":"Lin Zhu , Jiahui Hu , Luyi Bai","doi":"10.1016/j.knosys.2026.115387","DOIUrl":"10.1016/j.knosys.2026.115387","url":null,"abstract":"<div><div>Advancements in artificial intelligence has markedly highlighted the importance of temporal knowledge graphs. However, since factors such as limitations in data collection and the immaturity of knowledge extraction techniques, temporal knowledge graphs often remain incomplete. Moreover, the rapid growth of real-world information has significantly exacerbated the sparsity of these graphs, severely affecting their practical application effects. Precisely for these reasons, the enhancement of sparse temporal knowledge graphs has become a significant focus of investigation in current academic research. In the context of completing sparse temporal knowledge graphs, previous methods have enriched entity representations through neighbor information aggregation, alleviated the sparsity of the graphs, and improved the completion effect. However, these methods have limitations. On the one hand, they focus on the information of neighboring entities and neglect the role of the relationship vectors between the current entity and its neighbor entities. On the other hand, they fail to distinguish the aggregation weights according to the roles of adjacent entities, thus limiting the further improvement of the completion effect. Accordingly, this paper investigates an information aggregation method based on relevant facts and a role-oriented attention network to enrich entity representations. Given that the importance of relationship vectors is often overlooked, we propose a fact vector generation strategy through a chained relation extractor and a fact vector splicer to excavate the information of relationship vectors. Aiming at the problem that previous methods failed to distinguish the role weights of adjacent entities, we propose a role-oriented attention network. This network aggregates context information and assigns weights according to the roles of the aggregated information, thereby generating more accurate entity representations. According to the experimental results, our model outperforms state-of-the-art baseline models in the selected metrics.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"337 ","pages":"Article 115387"},"PeriodicalIF":7.6,"publicationDate":"2026-01-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146015907","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

SSTrack: Joint scale-aware temporal prompts and spatio-temporal prior transformer for visual object tracking 用于视觉目标跟踪的联合尺度感知时间提示和时空先验转换器

IF 7.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Knowledge-Based Systems

Pub Date : 2026-01-20 DOI: 10.1016/j.knosys.2026.115370

Sugang Ma , Zhen Wan , Bin Hu , Jinyu Zhang , Zhiqiang Hou , Xiangmo Zhao

Existing visual tracking algorithms have made impressive progress by leveraging the powerful global modeling capabilities of Transformers. However, these approaches typically focus on designing complex network models while neglecting temporal information and scale variations. These limitation makes them susceptible to tracking failures caused by target occlusion and deformation. Additionally, most trackers adopt ViT-based attention mechanisms. These trackers rely entirely on input images and lack task-relevant prior knowledge about the target. To address these issues, this paper proposes SSTrack, a novel visual tracking algorithm that integrates scale-aware temporal prompts and a spatio-temporal prior Transformer. Specifically, a scale-aware temporal information propagation mechanism is first designed, which allows the tracker to enable the model to learn the scale changes of the target between the preceding and following frames by propagating multi-scale temporal prompts across consecutive frames. Furthermore, we introduce a spatio-temporal prior module to provide the tracker with spatio-temporal prior knowledge of the target locations and appearances, combing spatio-temporal prior module with the self-attention module. Extensive experiments on seven benchmark datasets, including LaSOT, TrackingNet, and GOT-10k, demonstrate the superior tracking performance of SSTrack. The code and pre-trained models will be available at here.

现有的视觉跟踪算法通过利用变形金刚强大的全局建模能力取得了令人印象深刻的进展。然而，这些方法通常侧重于设计复杂的网络模型，而忽略了时间信息和尺度变化。这些限制使它们容易受到目标遮挡和变形引起的跟踪失败的影响。此外，大多数跟踪器采用基于虚拟现实的注意力机制。这些跟踪器完全依赖于输入图像，缺乏与目标任务相关的先验知识。为了解决这些问题，本文提出了一种新的视觉跟踪算法SSTrack，该算法集成了尺度感知时间提示和时空先验转换器。具体而言，首先设计了一种尺度感知的时间信息传播机制，该机制允许跟踪器通过跨连续帧传播多尺度时间提示，使模型能够了解目标在前帧和后帧之间的尺度变化。此外，我们将时空先验模块与自注意模块相结合，引入时空先验模块，为跟踪器提供目标位置和外观的时空先验知识。在LaSOT、TrackingNet和GOT-10k等7个基准数据集上进行的大量实验表明，SSTrack具有优越的跟踪性能。代码和预训练模型将在这里提供。

{"title":"SSTrack: Joint scale-aware temporal prompts and spatio-temporal prior transformer for visual object tracking","authors":"Sugang Ma , Zhen Wan , Bin Hu , Jinyu Zhang , Zhiqiang Hou , Xiangmo Zhao","doi":"10.1016/j.knosys.2026.115370","DOIUrl":"10.1016/j.knosys.2026.115370","url":null,"abstract":"<div><div>Existing visual tracking algorithms have made impressive progress by leveraging the powerful global modeling capabilities of Transformers. However, these approaches typically focus on designing complex network models while neglecting temporal information and scale variations. These limitation makes them susceptible to tracking failures caused by target occlusion and deformation. Additionally, most trackers adopt ViT-based attention mechanisms. These trackers rely entirely on input images and lack task-relevant prior knowledge about the target. To address these issues, this paper proposes SSTrack, a novel visual tracking algorithm that integrates scale-aware temporal prompts and a spatio-temporal prior Transformer. Specifically, a scale-aware temporal information propagation mechanism is first designed, which allows the tracker to enable the model to learn the scale changes of the target between the preceding and following frames by propagating multi-scale temporal prompts across consecutive frames. Furthermore, we introduce a spatio-temporal prior module to provide the tracker with spatio-temporal prior knowledge of the target locations and appearances, combing spatio-temporal prior module with the self-attention module. Extensive experiments on seven benchmark datasets, including LaSOT, TrackingNet, and GOT-10k, demonstrate the superior tracking performance of SSTrack. The code and pre-trained models will be available at <span><span>here</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"337 ","pages":"Article 115370"},"PeriodicalIF":7.6,"publicationDate":"2026-01-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146015909","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Learning deformable image registration with dilated attention transformer 用扩张注意力转换器学习变形图像配准

IF 7.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Knowledge-Based Systems

Pub Date : 2026-01-20 DOI: 10.1016/j.knosys.2026.115372

Yungeng Zhang , Yuan Chang , Xiaohou Shi , Yaqi Song , Ke Li , Feng Wang , Mingchuan Yang

Deformable image registration is a fundamental preprocessing step for many applications of medical image analysis. Recently, Transformers have demonstrated potential in deformable image registration. Transformers have the advantage of capturing spatial correlations within or across images by computing pairwise patch relations. However, due to the quadratic computational complexity of Transformers with respect to the sequence length and the fact that volumetric images contain an excessive number of voxels, current Transformer-based registration methods employ two strategies to mitigate the computational cost of Transformers when processing volumetric images. One approach is to integrate Transformers into the bottleneck of a CNN backbone to enhance low-resolution features, neglecting to use Transformers to capture high-resolution anatomical structure correlations. Another approach is to limit self-attention operations to small local windows, thereby restricting the receptive field of Transformers and their ability to deal with large displacements. Moreover, recent methods typically leverage the attention mechanism to enhance feature learning, without matching these features to explicitly calculate the displacement field. In this paper, we separate the processes of feature extraction, feature enhancement, and feature matching in deformable image registration. In order to process high-resolution 3D feature maps, we propose Dilated Attention Transformers. The Dilated Attention Transformers capture intra- and cross-image feature correlations with large receptive fields while ensuring manageable computational costs. Furthermore, in the feature matching process, we introduce a Dilated Matching strategy to accommodate large deformations. Experiments on public brain MRI and liver CT datasets demonstrate that our method performs favorably against the state-of-the-art deformable image registration methods.

形变图像配准是许多医学图像分析应用的基本预处理步骤。最近，变形金刚已经展示了变形图像配准的潜力。变压器具有通过计算成对补丁关系来捕获图像内部或图像之间的空间相关性的优点。然而，由于变形金刚相对于序列长度的二次计算复杂度和体积图像包含过多体素的事实，目前基于变形金刚的配准方法在处理体积图像时采用两种策略来降低变形金刚的计算成本。一种方法是将变形金刚集成到CNN主干的瓶颈中以增强低分辨率特征，而忽略了使用变形金刚来捕获高分辨率解剖结构相关性。另一种方法是将自我注意操作限制在小的局部窗口，从而限制变形金刚的接受范围及其处理大位移的能力。此外，最近的方法通常利用注意机制来增强特征学习，而不匹配这些特征来显式计算位移场。本文将可变形图像配准中的特征提取、特征增强和特征匹配过程分离开来。为了处理高分辨率的3D特征图，我们提出了扩展注意力转换器。扩张型注意力转换器在确保可管理的计算成本的同时，捕获具有大接受域的图像内部和跨图像特征相关性。此外，在特征匹配过程中，我们引入了一种适应大变形的扩展匹配策略。在公共脑MRI和肝脏CT数据集上的实验表明，我们的方法优于最先进的可变形图像配准方法。

{"title":"Learning deformable image registration with dilated attention transformer","authors":"Yungeng Zhang , Yuan Chang , Xiaohou Shi , Yaqi Song , Ke Li , Feng Wang , Mingchuan Yang","doi":"10.1016/j.knosys.2026.115372","DOIUrl":"10.1016/j.knosys.2026.115372","url":null,"abstract":"<div><div>Deformable image registration is a fundamental preprocessing step for many applications of medical image analysis. Recently, Transformers have demonstrated potential in deformable image registration. Transformers have the advantage of capturing spatial correlations within or across images by computing pairwise patch relations. However, due to the quadratic computational complexity of Transformers with respect to the sequence length and the fact that volumetric images contain an excessive number of voxels, current Transformer-based registration methods employ two strategies to mitigate the computational cost of Transformers when processing volumetric images. One approach is to integrate Transformers into the bottleneck of a CNN backbone to enhance low-resolution features, neglecting to use Transformers to capture high-resolution anatomical structure correlations. Another approach is to limit self-attention operations to small local windows, thereby restricting the receptive field of Transformers and their ability to deal with large displacements. Moreover, recent methods typically leverage the attention mechanism to enhance feature learning, without matching these features to explicitly calculate the displacement field. In this paper, we separate the processes of feature extraction, feature enhancement, and feature matching in deformable image registration. In order to process high-resolution 3D feature maps, we propose Dilated Attention Transformers. The Dilated Attention Transformers capture intra- and cross-image feature correlations with large receptive fields while ensuring manageable computational costs. Furthermore, in the feature matching process, we introduce a Dilated Matching strategy to accommodate large deformations. Experiments on public brain MRI and liver CT datasets demonstrate that our method performs favorably against the state-of-the-art deformable image registration methods.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"337 ","pages":"Article 115372"},"PeriodicalIF":7.6,"publicationDate":"2026-01-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146025751","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A temporal interaction-enhanced multi-view framework for knowledge tracing via self-supervised contrastive learning 基于自监督对比学习的时间交互增强多视角知识追踪框架

IF 7.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Knowledge-Based Systems

Pub Date : 2026-01-20 DOI: 10.1016/j.knosys.2026.115332

Liqing Qiu, Qingyun Zhu

Knowledge tracing, a fundamental technology in intelligent education systems, aims to analyze students’ historical question–answer records to infer their knowledge mastery and predict future learning performance. While existing graph-based approaches effectively capture complex dependencies between students and knowledge concepts, they remain limited in their ability to integrate heterogeneous information and model temporal dynamics. To address these challenges, this paper proposes a Temporal Interaction-Enhanced Multi-View Framework (TMCL) that leverages self-supervised contrastive learning. TMCL jointly models heterogeneous graphs and interaction-directed graphs to comprehensively integrate multi-dimensional static relationships and dynamic temporal dependencies among students, exercises, and skills. Furthermore, a self-supervised contrastive learning mechanism is incorporated to enhance the quality of heterogeneous node embeddings, while time-difference decay weights and gated temporal features are employed to improve the temporal sensitivity of interaction graph modeling. Experimental results on three publicly available datasets demonstrate that TMCL achieves consistent performance gains across different datasets, with maximum improvements of 3.10% in AUC and 2.59% in ACC, thereby validating its effectiveness and advantages in personalized learning prediction tasks.

知识溯源是智能教育系统的一项基础技术，其目的是通过分析学生的历史问答记录，推断学生对知识的掌握程度，预测未来的学习表现。虽然现有的基于图的方法有效地捕获了学生和知识概念之间的复杂依赖关系，但它们在集成异构信息和建模时间动态方面的能力仍然有限。为了解决这些挑战，本文提出了一种利用自监督对比学习的时间交互增强多视图框架（TMCL）。TMCL联合建模异构图和面向交互的图，全面集成学生、练习和技能之间的多维静态关系和动态时间依赖关系。此外，采用自监督对比学习机制提高异构节点嵌入质量，采用时差衰减权值和门控时间特征提高交互图建模的时间敏感性。在三个公开的数据集上的实验结果表明，TMCL在不同的数据集上取得了一致的性能提升，AUC和ACC的最大提升分别为3.10%和2.59%，从而验证了其在个性化学习预测任务中的有效性和优势。

{"title":"A temporal interaction-enhanced multi-view framework for knowledge tracing via self-supervised contrastive learning","authors":"Liqing Qiu, Qingyun Zhu","doi":"10.1016/j.knosys.2026.115332","DOIUrl":"10.1016/j.knosys.2026.115332","url":null,"abstract":"<div><div>Knowledge tracing, a fundamental technology in intelligent education systems, aims to analyze students’ historical question–answer records to infer their knowledge mastery and predict future learning performance. While existing graph-based approaches effectively capture complex dependencies between students and knowledge concepts, they remain limited in their ability to integrate heterogeneous information and model temporal dynamics. To address these challenges, this paper proposes a Temporal Interaction-Enhanced Multi-View Framework (TMCL) that leverages self-supervised contrastive learning. TMCL jointly models heterogeneous graphs and interaction-directed graphs to comprehensively integrate multi-dimensional static relationships and dynamic temporal dependencies among students, exercises, and skills. Furthermore, a self-supervised contrastive learning mechanism is incorporated to enhance the quality of heterogeneous node embeddings, while time-difference decay weights and gated temporal features are employed to improve the temporal sensitivity of interaction graph modeling. Experimental results on three publicly available datasets demonstrate that TMCL achieves consistent performance gains across different datasets, with maximum improvements of 3.10% in AUC and 2.59% in ACC, thereby validating its effectiveness and advantages in personalized learning prediction tasks.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"337 ","pages":"Article 115332"},"PeriodicalIF":7.6,"publicationDate":"2026-01-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146015905","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Constructing trend-based information granules for enhanced long-term time series analysis and forecasting 构建基于趋势的信息颗粒，增强长期时间序列分析和预测

IF 7.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Knowledge-Based Systems

Pub Date : 2026-01-19 DOI: 10.1016/j.knosys.2026.115381

Mingli Song , Qingyu Wang , Haoyang Yu , Witold Pedrycz

This study presents an approach to construct a set of trend-based quintuple information granules on numeric time series and apply the information granules to make long-term forecasting. The set of information granules not only capture local features and boundary information more intuitively, but also provides more useful knowledge for future predictions, thereby improving the robustness of the predictions. There are three key innovations. First, a dynamic trend-based granulation approach and the corresponding evaluation metric are proposed, in which the principle of justifiable granularity is integrated with trend segmentation. By adaptively optimizing data granularity through volatility and trend consistency analysis, it overcomes fixed-window limitations and improves feature representation. Second, a hybrid optimization algorithm is developed by combining the genetic algorithm with particle swarm optimization, enhancing global search efficiency and segmentation precision. A dynamic inertia weight adjustment mechanism further stabilizes segmentation results. Third, a hybrid Transformer-LSTM model is designed to process information granules, uniting LSTM’s strength in capturing long-term dependencies with the Transformer’s global feature extraction through multihead attention and positional encoding. This integration improves both prediction accuracy and temporal pattern recognition. Collectively, these components address major challenges in long-term forecasting, particularly managing complex temporal dynamics and ensuring consistency over extended horizons. In interval forecasting analysis, our information granulation approach can achieve at least 80% coverage for all datasets using more compact granules.

本文提出了在数值时间序列上构造一组基于趋势的五元信息粒，并应用该信息粒进行长期预测的方法。信息颗粒集不仅可以更直观地捕获局部特征和边界信息，还可以为未来的预测提供更多有用的知识，从而提高预测的鲁棒性。有三个关键的创新。首先，将合理粒度原则与趋势分割相结合，提出了一种基于趋势的动态颗粒化方法及其评价指标；通过波动性和趋势一致性分析自适应优化数据粒度，克服了固定窗口的限制，改进了特征表示。其次，将遗传算法与粒子群算法相结合，提出了一种混合优化算法，提高了全局搜索效率和分割精度；动态惯性权重调整机制进一步稳定了分割结果。第三，设计了一种混合Transformer-LSTM模型来处理信息颗粒，将LSTM在捕获长期依赖关系方面的优势与Transformer通过多头注意和位置编码的全局特征提取相结合。这种集成提高了预测精度和时间模式识别。总的来说，这些组成部分解决了长期预测中的主要挑战，特别是管理复杂的时间动态和确保延长视界的一致性。在区间预测分析中，我们的信息粒化方法可以使用更紧凑的颗粒对所有数据集实现至少80%的覆盖率。

{"title":"Constructing trend-based information granules for enhanced long-term time series analysis and forecasting","authors":"Mingli Song , Qingyu Wang , Haoyang Yu , Witold Pedrycz","doi":"10.1016/j.knosys.2026.115381","DOIUrl":"10.1016/j.knosys.2026.115381","url":null,"abstract":"<div><div>This study presents an approach to construct a set of trend-based quintuple information granules on numeric time series and apply the information granules to make long-term forecasting. The set of information granules not only capture local features and boundary information more intuitively, but also provides more useful knowledge for future predictions, thereby improving the robustness of the predictions. There are three key innovations. First, a dynamic trend-based granulation approach and the corresponding evaluation metric are proposed, in which the principle of justifiable granularity is integrated with trend segmentation. By adaptively optimizing data granularity through volatility and trend consistency analysis, it overcomes fixed-window limitations and improves feature representation. Second, a hybrid optimization algorithm is developed by combining the genetic algorithm with particle swarm optimization, enhancing global search efficiency and segmentation precision. A dynamic inertia weight adjustment mechanism further stabilizes segmentation results. Third, a hybrid Transformer-LSTM model is designed to process information granules, uniting LSTM’s strength in capturing long-term dependencies with the Transformer’s global feature extraction through multihead attention and positional encoding. This integration improves both prediction accuracy and temporal pattern recognition. Collectively, these components address major challenges in long-term forecasting, particularly managing complex temporal dynamics and ensuring consistency over extended horizons. In interval forecasting analysis, our information granulation approach can achieve at least 80% coverage for all datasets using more compact granules.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"336 ","pages":"Article 115381"},"PeriodicalIF":7.6,"publicationDate":"2026-01-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146024142","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Multi-contrast feature cross entanglement network for joint MR image reconstruction and super-resolution 多对比度特征交叉纠缠网络联合磁共振图像重建和超分辨率

IF 7.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Knowledge-Based Systems

Pub Date : 2026-01-19 DOI: 10.1016/j.knosys.2026.115368

Guoqing Ge , Weisheng Li , Yucheng Shu , Xiaoyu Qiao

Reconstruction and super-resolution (SR) provide effective solutions for accelerating multi-contrast magnetic resonance (MR) imaging by leveraging auxiliary contrast information to restore target contrast from an undersampled counterpart. Although recent advances have explored the joint optimization of reconstruction and SR, most existing frameworks still adopt shallow concatenation or independent decoding branches, thereby failing to fully exploit the inherent complementarity and hierarchical correlations between the two tasks. Additionally, auxiliary contrast information is typically integrated in an isotropic and coarse-grained manner, neglecting directional and structure-specific dependencies across anatomical regions, thus weakening its ability to provide discriminative guidance for the target contrast reconstruction. To address these limitations, we propose a multi-contrast feature cross entanglement network (MFCE-Net) that facilitates comprehensive feature interaction across modalities and tasks. In detail, we first introduce a multi-branch feature guidance module to facilitate multi-scale and direction-aware feature transfer across modalities. Furthermore, within the designed top-down architecture, we incorporate an attention mechanism that allows the SR branch to capture global structures while preserving fine textures by proposing a feature representation enhancement module. Finally, we design a feature entanglement interaction (FEI) module that employs a cross-weighting mechanism across spatial and channel dimensions to facilitate deep feature sharing and mutual reinforcement between the reconstruction and SR tasks. Extensive experiments are conducted with various advanced multi-contrast MR imaging methods on fastMRI, BraTS2019 and clinical in-house datasets, and the results demonstrate the superiority of our model. The code is released at https://github.com/coolggq/MFCE-Net.

重建和超分辨率（SR）为加速多对比度磁共振（MR）成像提供了有效的解决方案，通过利用辅助对比度信息从欠采样对应物中恢复目标对比度。虽然近年来已经探索了重建和SR的联合优化，但大多数现有框架仍然采用浅连接或独立解码分支，从而未能充分利用这两个任务之间固有的互补性和层次相关性。此外，辅助对比度信息通常以各向同性和粗粒度的方式集成，忽略了跨解剖区域的方向性和结构特异性依赖，从而削弱了其为目标对比度重建提供鉴别指导的能力。为了解决这些限制，我们提出了一个多对比特征交叉纠缠网络（MFCE-Net），促进了跨模式和任务的综合特征交互。我们首先介绍了一个多分支特征引导模块，以促进多尺度和方向感知的特征跨模态传输。此外，在设计的自顶向下架构中，我们结合了一个注意机制，通过提出一个特征表示增强模块，允许SR分支捕获全局结构，同时保留精细纹理。最后，我们设计了一个特征纠缠交互（FEI）模块，该模块采用跨空间和通道维度的交叉加权机制，以促进重建和SR任务之间的深度特征共享和相互增强。在fastMRI、BraTS2019和临床内部数据集上使用各种先进的多对比磁共振成像方法进行了大量实验，结果证明了我们模型的优越性。该代码发布在https://github.com/coolggq/MFCE-Net。

{"title":"Multi-contrast feature cross entanglement network for joint MR image reconstruction and super-resolution","authors":"Guoqing Ge , Weisheng Li , Yucheng Shu , Xiaoyu Qiao","doi":"10.1016/j.knosys.2026.115368","DOIUrl":"10.1016/j.knosys.2026.115368","url":null,"abstract":"<div><div>Reconstruction and super-resolution (SR) provide effective solutions for accelerating multi-contrast magnetic resonance (MR) imaging by leveraging auxiliary contrast information to restore target contrast from an undersampled counterpart. Although recent advances have explored the joint optimization of reconstruction and SR, most existing frameworks still adopt shallow concatenation or independent decoding branches, thereby failing to fully exploit the inherent complementarity and hierarchical correlations between the two tasks. Additionally, auxiliary contrast information is typically integrated in an isotropic and coarse-grained manner, neglecting directional and structure-specific dependencies across anatomical regions, thus weakening its ability to provide discriminative guidance for the target contrast reconstruction. To address these limitations, we propose a multi-contrast feature cross entanglement network (MFCE-Net) that facilitates comprehensive feature interaction across modalities and tasks. In detail, we first introduce a multi-branch feature guidance module to facilitate multi-scale and direction-aware feature transfer across modalities. Furthermore, within the designed top-down architecture, we incorporate an attention mechanism that allows the SR branch to capture global structures while preserving fine textures by proposing a feature representation enhancement module. Finally, we design a feature entanglement interaction (FEI) module that employs a cross-weighting mechanism across spatial and channel dimensions to facilitate deep feature sharing and mutual reinforcement between the reconstruction and SR tasks. Extensive experiments are conducted with various advanced multi-contrast MR imaging methods on fastMRI, BraTS2019 and clinical in-house datasets, and the results demonstrate the superiority of our model. The code is released at <span><span>https://github.com/coolggq/MFCE-Net</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"337 ","pages":"Article 115368"},"PeriodicalIF":7.6,"publicationDate":"2026-01-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080655","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Fontify : One-shot font generation via in-context learning Fontify：通过上下文学习一次性生成字体

IF 7.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Knowledge-Based Systems

Pub Date : 2026-01-18 DOI: 10.1016/j.knosys.2026.115354

Ying Xu , Xiangwei Zhu , Songyuan Li

Automatic font generation has attracted increasing interest due to the demand for rapid creation of stylized glyph libraries. However, existing methods often depend on predefined style priors, fixed style embeddings, or strictly aligned reference sets, which limit their ability to generalize to unseen styles-particularly in one-shot settings. The core difficulty in one-shot font generation lies in the ambiguity of aligning content strokes with stylistic patterns, which makes it challenging to synthesize coherent glyphs from a single reference. To address these limitations, we propose Fontify, a one-shot font generation framework that reformulates glyph synthesis as a context-aware image inpainting task. By concatenating content–style pairs and applying random block-wise masking during training, the model learns to infer missing stylistic details from partial contextual cues. At inference, Fontify generates high-fidelity glyphs through visual prompting, requiring no predefined font priors or explicit style embeddings. Extensive experiments demonstrate that Fontify outperforms existing approaches on both seen and unseen style scenarios, producing glyphs with higher stroke fidelity, sharper structural consistency, and improved perceptual realism. Our work introduces a data-efficient paradigm for font generation with potential applications in digital typography, artistic design, and personalized font creation. The code is available at https://github.com/YingXu124/Fontify/.

由于需要快速创建风格化的字形库，自动字体生成引起了越来越多的兴趣。然而，现有的方法通常依赖于预定义的样式先验、固定的样式嵌入或严格对齐的参考集，这限制了它们泛化到不可见样式的能力——特别是在一次性设置中。一次性字体生成的核心困难在于内容笔画与风格模式对齐的模糊性，这使得从单个引用中合成连贯的字形具有挑战性。为了解决这些限制，我们提出了Fontify，这是一个一次性字体生成框架，它将字形合成重新制定为上下文感知的图像绘制任务。通过连接内容风格对并在训练期间应用随机块屏蔽，该模型学会从部分上下文线索中推断缺失的风格细节。在推理时，Fontify通过视觉提示生成高保真的字形，不需要预定义的字体先验或显式的样式嵌入。大量的实验表明，Fontify在可见和不可见的风格场景上都优于现有的方法，产生的字形具有更高的笔画保真度、更清晰的结构一致性和改进的感知真实感。我们的工作介绍了一种数据高效的字体生成范例，在数字排版、艺术设计和个性化字体创建方面具有潜在的应用。代码可在https://github.com/YingXu124/Fontify/上获得。

{"title":"Fontify : One-shot font generation via in-context learning","authors":"Ying Xu , Xiangwei Zhu , Songyuan Li","doi":"10.1016/j.knosys.2026.115354","DOIUrl":"10.1016/j.knosys.2026.115354","url":null,"abstract":"<div><div>Automatic font generation has attracted increasing interest due to the demand for rapid creation of stylized glyph libraries. However, existing methods often depend on predefined style priors, fixed style embeddings, or strictly aligned reference sets, which limit their ability to generalize to unseen styles-particularly in one-shot settings. The core difficulty in one-shot font generation lies in the ambiguity of aligning content strokes with stylistic patterns, which makes it challenging to synthesize coherent glyphs from a single reference. To address these limitations, we propose Fontify, a one-shot font generation framework that reformulates glyph synthesis as a context-aware image inpainting task. By concatenating content–style pairs and applying random block-wise masking during training, the model learns to infer missing stylistic details from partial contextual cues. At inference, Fontify generates high-fidelity glyphs through visual prompting, requiring no predefined font priors or explicit style embeddings. Extensive experiments demonstrate that Fontify outperforms existing approaches on both seen and unseen style scenarios, producing glyphs with higher stroke fidelity, sharper structural consistency, and improved perceptual realism. Our work introduces a data-efficient paradigm for font generation with potential applications in digital typography, artistic design, and personalized font creation. The code is available at <span><span>https://github.com/YingXu124/Fontify/</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"336 ","pages":"Article 115354"},"PeriodicalIF":7.6,"publicationDate":"2026-01-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146024315","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

MEU-miner: Bitset-guided multi-dimensional high-utility sequential pattern mining with parallelism and adaptive top-k MEU-miner: bitset引导的具有并行性和自适应top-k的多维高效顺序模式挖掘

IF 7.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Knowledge-Based Systems

Pub Date : 2026-01-17 DOI: 10.1016/j.knosys.2026.115314

Wenhao Zhao , Cewen Tian , Xingyu Wang , Dezhen Wang , Wei Zhang

High-utility sequential pattern mining (HUSPM) aims to uncover valuable sequential behaviors to support decision making across domains such as e-commerce analytics and Web interaction modeling. Despite extensions to multi-dimensional settings, existing methods remain limited by context-agnostic pruning that inflates candidate sets and degrades runtime performance, list-based utility storage whose memory footprint balloons as dimensionality grows, and the absence of parallel execution and adaptive Top-k control, which together constrain scalability on large datasets. To address these challenges, we propose a family of methods-MEU-Miner, PMEU-Miner, and TKMEU-Miner-for scalable multi-dimensional HUSPM (MDHUSPM) on commodity hardware. MEU-Miner couples bitset-driven context matching with a dimension-level utility bound (DUB) to enable safe pruning and supports incremental updates/caching of matching vectors and bounds; DUB is a safe upper bound guaranteed by the anti-monotonicity of the context lattice, ensuring that no valid multi-dimensional high-utility pattern is ever pruned. PMEU-Miner adopts a shared-memory parallel design (thread pool with work stealing and fine-grained critical sections) to obtain substantial multi-core speedups; and TKMEU-Miner provides threshold-free Top-k mining via an adaptive heap with bound-guided early termination. Our experiments span 19 datasets, including 13 real-world and 6 synthetic databases. Extensive experiments on 19 datasets show that MEU-Miner achieves up to 66 ×  speedups and reduces candidates by up to 80% over state-of-the-art MDHUSPM baselines; PMEU-Miner accelerates mining by up to 180 ×  through multi-threaded search; and TKMEU-Miner outperforms Top-k competitors by 10 ×  to 40 ×  across different values of k. Our framework also maintains stable memory usage as dimensionality increases, advancing the state of the art and enabling fast, reproducible pattern discovery in context-rich sequential data.

高实用顺序模式挖掘（High-utility sequential pattern mining， HUSPM）旨在揭示有价值的顺序行为，以支持跨领域（如电子商务分析和Web交互建模）的决策制定。尽管对多维设置进行了扩展，但现有方法仍然受到以下方面的限制：与上下文无关的修剪会使候选集膨胀并降低运行时性能；基于列表的实用程序存储会随着维度的增长而占用内存；缺乏并行执行和自适应Top-k控制，这些因素共同限制了大型数据集的可伸缩性。为了应对这些挑战，我们提出了一系列方法- meu - miner， PMEU-Miner和tkmeu - miner -用于商用硬件上的可扩展多维HUSPM （MDHUSPM）。MEU-Miner将位集驱动的上下文匹配与维度级实用程序绑定（DUB）相结合，以实现安全修剪，并支持匹配向量和边界的增量更新/缓存；DUB是由上下文格的反单调性保证的安全上界，保证了有效的多维高效用模式不被剪枝。PMEU-Miner采用共享内存并行设计（线程池与工作窃取和细粒度临界区），以获得大量的多核速度；TKMEU-Miner通过具有边界引导的早期终止的自适应堆提供无阈值的Top-k挖掘。我们的实验跨越19个数据集，包括13个真实数据库和6个合成数据库。在19个数据集上进行的广泛实验表明，MEU-Miner实现了高达66 × 的加速，并比最先进的MDHUSPM基线减少了80%的候选数据；PMEU-Miner通过多线程搜索加速挖掘高达180 × ；在不同的k值上，TKMEU-Miner比Top-k的竞争对手高出10 × 到40 × 。随着维数的增加，我们的框架还保持了稳定的内存使用，推进了最先进的技术，并在上下文丰富的顺序数据中实现了快速、可重复的模式发现。

{"title":"MEU-miner: Bitset-guided multi-dimensional high-utility sequential pattern mining with parallelism and adaptive top-k","authors":"Wenhao Zhao , Cewen Tian , Xingyu Wang , Dezhen Wang , Wei Zhang","doi":"10.1016/j.knosys.2026.115314","DOIUrl":"10.1016/j.knosys.2026.115314","url":null,"abstract":"<div><div>High-utility sequential pattern mining (HUSPM) aims to uncover valuable sequential behaviors to support decision making across domains such as e-commerce analytics and Web interaction modeling. Despite extensions to multi-dimensional settings, existing methods remain limited by context-agnostic pruning that inflates candidate sets and degrades runtime performance, list-based utility storage whose memory footprint balloons as dimensionality grows, and the absence of parallel execution and adaptive Top-k control, which together constrain scalability on large datasets. To address these challenges, we propose a family of methods-MEU-Miner, PMEU-Miner, and TKMEU-Miner-for scalable multi-dimensional HUSPM (MDHUSPM) on commodity hardware. MEU-Miner couples bitset-driven context matching with a dimension-level utility bound (DUB) to enable safe pruning and supports incremental updates/caching of matching vectors and bounds; DUB is a safe upper bound guaranteed by the anti-monotonicity of the context lattice, ensuring that no valid multi-dimensional high-utility pattern is ever pruned. PMEU-Miner adopts a shared-memory parallel design (thread pool with work stealing and fine-grained critical sections) to obtain substantial multi-core speedups; and TKMEU-Miner provides threshold-free Top-k mining via an adaptive heap with bound-guided early termination. Our experiments span 19 datasets, including 13 real-world and 6 synthetic databases. Extensive experiments on 19 datasets show that MEU-Miner achieves up to 66 ×  speedups and reduces candidates by up to 80% over state-of-the-art MDHUSPM baselines; PMEU-Miner accelerates mining by up to 180 ×  through multi-threaded search; and TKMEU-Miner outperforms Top-<em>k</em> competitors by 10 ×  to 40 ×  across different values of <em>k</em>. Our framework also maintains stable memory usage as dimensionality increases, advancing the state of the art and enabling fast, reproducible pattern discovery in context-rich sequential data.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"337 ","pages":"Article 115314"},"PeriodicalIF":7.6,"publicationDate":"2026-01-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146081105","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0