Information Sciences最新文献_第10页

Reducing contextual noise in review-based recommendation via aspect term extraction and attention modeling 通过方面词提取和注意力建模减少基于评论的推荐中的上下文噪声

IF 6.8 1区计算机科学 0 COMPUTER SCIENCE, INFORMATION SYSTEMS

Information Sciences

Pub Date : 2026-01-07 DOI: 10.1016/j.ins.2026.123078

Heena Lim , Xinzhe Li , Seonu Park , Qinglong Li , Jaekyeong Kim

Information and Communication Technology (ICT) advances have accelerated e-commerce growth, exposing users to an overwhelming number of products and producing information overload. Recommender systems mitigate this problem by modeling user–item interactions, but traditional matrix factorization (MF) methods suffer from data sparsity. Prior work leverages user-generated review text to supply semantic cues about preferences and product attributes. However, most methods process whole reviews indiscriminately, mixing aspect-relevant content with contextual noise. This noise reduces the informational density, defined as the proportion of aspect-relevant content in a document, thereby degrading textual representations and model accuracy. We propose the Aspect Term-aware Recommender System (ATRS), which incorporates aspect-level semantics into review-based recommendation. ATRS applies a Bidirectional Encoder Representations from Transformers (BERT)-based aspect term extraction (ATE) model to identify product-related terms and filter irrelevant content, increasing informational density. Extracted aspect terms are encoded by a convolutional neural network (CNN) and aggregated with self-attention to produce aspect-aware user and item representations. Experiments on Amazon and Yelp datasets show ATRS outperforms baselines, producing average improvements of 19.54% in mean absolute error (MAE) and 11.89% in root mean square error (RMSE). Results confirm the benefit of aspect-level refinement and optimizing informational density in review-based recommender systems.

信息和通信技术（ICT）的进步加速了电子商务的发展，使用户接触到大量的产品，产生了信息过载。推荐系统通过对用户与物品的交互进行建模来缓解这一问题，但传统的矩阵分解（MF）方法存在数据稀疏性的问题。先前的工作利用用户生成的评论文本来提供关于偏好和产品属性的语义线索。然而，大多数方法不加选择地处理整个评论，将与方面相关的内容与上下文噪声混合在一起。这种噪声降低了信息密度（定义为文档中与方面相关的内容的比例），从而降低了文本表示和模型准确性。我们提出了面向方面的术语感知推荐系统（ATRS），该系统将面向方面的语义整合到基于评论的推荐中。ATRS应用基于BERT（双向编码器表示）的方面术语提取（ATE）模型来识别与产品相关的术语并过滤不相关的内容，从而增加信息密度。提取的方面术语由卷积神经网络（CNN）编码，并与自关注聚合以产生方面感知的用户和项目表示。在亚马逊和Yelp数据集上的实验表明，ATRS优于基线，平均绝对误差（MAE）和均方根误差（RMSE）的平均改进分别为19.54%和11.89%。结果证实了在基于评论的推荐系统中，方面级精化和优化信息密度的好处。

{"title":"Reducing contextual noise in review-based recommendation via aspect term extraction and attention modeling","authors":"Heena Lim , Xinzhe Li , Seonu Park , Qinglong Li , Jaekyeong Kim","doi":"10.1016/j.ins.2026.123078","DOIUrl":"10.1016/j.ins.2026.123078","url":null,"abstract":"<div><div>Information and Communication Technology (ICT) advances have accelerated e-commerce growth, exposing users to an overwhelming number of products and producing information overload. Recommender systems mitigate this problem by modeling user–item interactions, but traditional matrix factorization (MF) methods suffer from data sparsity. Prior work leverages user-generated review text to supply semantic cues about preferences and product attributes. However, most methods process whole reviews indiscriminately, mixing aspect-relevant content with contextual noise. This noise reduces the informational density, defined as the proportion of aspect-relevant content in a document, thereby degrading textual representations and model accuracy. We propose the Aspect Term-aware Recommender System (ATRS), which incorporates aspect-level semantics into review-based recommendation. ATRS applies a Bidirectional Encoder Representations from Transformers (BERT)-based aspect term extraction (ATE) model to identify product-related terms and filter irrelevant content, increasing informational density. Extracted aspect terms are encoded by a convolutional neural network (CNN) and aggregated with self-attention to produce aspect-aware user and item representations. Experiments on Amazon and Yelp datasets show ATRS outperforms baselines, producing average improvements of 19.54% in mean absolute error (MAE) and 11.89% in root mean square error (RMSE). Results confirm the benefit of aspect-level refinement and optimizing informational density in review-based recommender systems.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"735 ","pages":"Article 123078"},"PeriodicalIF":6.8,"publicationDate":"2026-01-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145940331","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-scale skin lesion segmentation network with dynamic mask guidance and dual-path adaptive attention 基于动态掩模引导和双路径自适应关注的多尺度皮肤病变分割网络

IF 6.8 1区计算机科学 0 COMPUTER SCIENCE, INFORMATION SYSTEMS

Information Sciences

Pub Date : 2026-01-06 DOI: 10.1016/j.ins.2026.123083

Hao Liu, Yachao Li, Jiuzhen Liang

Precise skin lesion segmentation is crucial for clinical diagnosis and treatment, yet remains challenging due to heterogeneous shapes, blurred boundaries, and low-contrast. Existing encoder-decoder methods rely on static masks and struggle to capture fine boundaries of small lesions. Moreover, simple cross-layer feature fusion is insufficient to alleviate semantic inconsistencies across scales. These limitations reduce segmentation accuracy in low-contrast regions and small lesions. To address these issues, this study proposes MS-DMGNet, a multi-scale segmentation framework incorporating dynamic mask guidance and dual-path adaptive attention. The Dynamic Mask-Guided Decoder (DMGD) treats the predicted mask as an adaptive spatial signal. This signal actively guides multi-scale feature propagation and refinement, leading to more precise boundary recovery. The Dual-Path Adaptive Attention Mechanism (DPAM) enhances cross-level semantic consistency by aligning fine spatial details with high-level contextual cues, effectively mitigating semantic drift. In addition, the Hierarchical Loss Fusion (HLF) strategy supervises feature optimization at each decoding stage, promoting cooperative refinement across the network. Experiments on ISIC2016, ISIC2017, ISIC2018, and CVC-ClinicDB show Dice scores of

93.04 %

,

91.23 %

,

90.82 %

, and

92.98 %

, respectively, while cross-dataset transfer from ISIC2018 to PH2 highlights strong generalization. Overall, MS-DMGNet achieves robust performance and demonstrates promising clinical applicability.

精确的皮肤病变分割对临床诊断和治疗至关重要，但由于形状不均匀、边界模糊和对比度低，仍然具有挑战性。现有的编码器-解码器方法依赖于静态掩模，难以捕捉小病灶的精细边界。此外，简单的跨层特征融合不足以缓解跨尺度的语义不一致。这些限制降低了低对比度区域和小病变的分割准确性。为了解决这些问题，本研究提出了MS-DMGNet，一种结合动态掩模引导和双路径自适应关注的多尺度分割框架。动态掩码引导解码器（DMGD）将预测的掩码作为自适应空间信号处理。该信号主动引导多尺度特征传播和细化，从而更精确地恢复边界。双路径自适应注意机制（Dual-Path Adaptive Attention Mechanism， DPAM）通过将精细的空间细节与高水平的上下文线索对齐来增强跨层语义一致性，有效地缓解了语义漂移。此外，层次损失融合（HLF）策略监督每个解码阶段的特征优化，促进整个网络的协同细化。在ISIC2016、ISIC2017、ISIC2018和CVC-ClinicDB上的实验显示，Dice得分分别为93.04%、91.23%、90.82%和92.98%，而从ISIC2018到PH2的跨数据集迁移显示出较强的泛化能力。总体而言，MS-DMGNet实现了稳健的性能，并显示出有希望的临床适用性。

{"title":"Multi-scale skin lesion segmentation network with dynamic mask guidance and dual-path adaptive attention","authors":"Hao Liu, Yachao Li, Jiuzhen Liang","doi":"10.1016/j.ins.2026.123083","DOIUrl":"10.1016/j.ins.2026.123083","url":null,"abstract":"<div><div>Precise skin lesion segmentation is crucial for clinical diagnosis and treatment, yet remains challenging due to heterogeneous shapes, blurred boundaries, and low-contrast. Existing encoder-decoder methods rely on static masks and struggle to capture fine boundaries of small lesions. Moreover, simple cross-layer feature fusion is insufficient to alleviate semantic inconsistencies across scales. These limitations reduce segmentation accuracy in low-contrast regions and small lesions. To address these issues, this study proposes MS-DMGNet, a multi-scale segmentation framework incorporating dynamic mask guidance and dual-path adaptive attention. The Dynamic Mask-Guided Decoder (DMGD) treats the predicted mask as an adaptive spatial signal. This signal actively guides multi-scale feature propagation and refinement, leading to more precise boundary recovery. The Dual-Path Adaptive Attention Mechanism (DPAM) enhances cross-level semantic consistency by aligning fine spatial details with high-level contextual cues, effectively mitigating semantic drift. In addition, the Hierarchical Loss Fusion (HLF) strategy supervises feature optimization at each decoding stage, promoting cooperative refinement across the network. Experiments on ISIC2016, ISIC2017, ISIC2018, and CVC-ClinicDB show Dice scores of <span><math><mn>93.04</mn><mi>%</mi></math></span>, <span><math><mn>91.23</mn><mi>%</mi></math></span>, <span><math><mn>90.82</mn><mi>%</mi></math></span>, and <span><math><mn>92.98</mn><mi>%</mi></math></span>, respectively, while cross-dataset transfer from ISIC2018 to PH2 highlights strong generalization. Overall, MS-DMGNet achieves robust performance and demonstrates promising clinical applicability.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"735 ","pages":"Article 123083"},"PeriodicalIF":6.8,"publicationDate":"2026-01-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145940389","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

UniTrack: Unifying day and night tracking with continual learning UniTrack：统一昼夜跟踪和持续学习

IF 6.8 1区计算机科学 0 COMPUTER SCIENCE, INFORMATION SYSTEMS

Information Sciences

Pub Date : 2026-01-06 DOI: 10.1016/j.ins.2026.123081

Jiwei Mo , Feixiang He , Pengzhi Zhong , Qijun Zhao , Dan Zeng , Ling Huang , Shuiwang Li , Xianhao Shen

Illumination variation remains one of the most persistent challenges in single object tracking (SOT), as drastic changes between day and night often lead to significant performance degradation. To the best of our knowledge, this work is the first to explore a Continual Learning (CL) framework for SOT, aiming to tackle the challenges arising from diverse and dynamically changing illumination conditions. Leveraging the adaptability of prompting-based Vision Transformers (ViTs), our approach tackles distribution shifts that arise in both daytime and nighttime scenarios. Unlike existing prompting-based CL methods that treat prompts across ViT layers independently, we propose an adaptive prompt aggregation module that models hierarchical dependencies among prompts across transformer blocks. This design enables more robust adaptation to changing visual environments. Extensive experiments demonstrate that our method outperforms state-of-the-art trackers, with AUC gains of 1.88% on DarkTrack2021 and 3.8% on VisDrone2018. Source code is available at https://github.com/Gkk-Code/UniTrack.

光照变化仍然是单目标跟踪（SOT）中最持久的挑战之一，因为昼夜之间的剧烈变化通常会导致显著的性能下降。据我们所知，这项工作是第一个探索SOT的持续学习（CL）框架，旨在解决各种动态变化的照明条件所带来的挑战。利用基于提示的视觉变压器（ViTs）的适应性，我们的方法可以处理白天和夜间场景中出现的分布变化。与现有的基于提示的CL方法不同，我们提出了一个自适应的提示聚合模块，该模块对跨变压器块的提示之间的分层依赖关系进行建模。这种设计能够更好地适应不断变化的视觉环境。大量实验表明，我们的方法优于最先进的跟踪器，在DarkTrack2021和VisDrone2018上的AUC增益分别为1.88%和3.8%。源代码可从https://github.com/Gkk-Code/UniTrack获得。

{"title":"UniTrack: Unifying day and night tracking with continual learning","authors":"Jiwei Mo , Feixiang He , Pengzhi Zhong , Qijun Zhao , Dan Zeng , Ling Huang , Shuiwang Li , Xianhao Shen","doi":"10.1016/j.ins.2026.123081","DOIUrl":"10.1016/j.ins.2026.123081","url":null,"abstract":"<div><div>Illumination variation remains one of the most persistent challenges in single object tracking (SOT), as drastic changes between day and night often lead to significant performance degradation. To the best of our knowledge, this work is the first to explore a Continual Learning (CL) framework for SOT, aiming to tackle the challenges arising from diverse and dynamically changing illumination conditions. Leveraging the adaptability of prompting-based Vision Transformers (ViTs), our approach tackles distribution shifts that arise in both daytime and nighttime scenarios. Unlike existing prompting-based CL methods that treat prompts across ViT layers independently, we propose an adaptive prompt aggregation module that models hierarchical dependencies among prompts across transformer blocks. This design enables more robust adaptation to changing visual environments. Extensive experiments demonstrate that our method outperforms state-of-the-art trackers, with AUC gains of 1.88% on DarkTrack2021 and 3.8% on VisDrone2018. Source code is available at <span><span>https://github.com/Gkk-Code/UniTrack</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"735 ","pages":"Article 123081"},"PeriodicalIF":6.8,"publicationDate":"2026-01-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145940392","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

Structural propagation dual-prototype refinement network for few-shot 3D point cloud segmentation 基于结构传播双原型的少镜头三维点云分割方法

IF 6.8 1区计算机科学 0 COMPUTER SCIENCE, INFORMATION SYSTEMS

Information Sciences

Pub Date : 2026-01-06 DOI: 10.1016/j.ins.2026.123076

Qingzheng Wang , Jiazhi Xie , Ning Li, Jingjun Bi, Zengwei Mai, Wenhui Liu, Xingqin Wang

Few-shot 3D point cloud semantic segmentation (FS-PCS) aims to recognize novel classes with only limited annotated examples. Existing methods face two key challenges. Sliding-window preprocessing of large scenes often truncates object instances and induces structural bias, while masked average pooling prototypes struggle to capture these fragmented structures. Furthermore, many methods prioritize query-space refined prototypes over support-derived base prototypes, leading to a potential compromise of stable, discriminative cues derived from the original support structure. To address these issues, we propose the Structural Propagation Dual-Prototype Refinement Network (SP-DRNet). A lightweight Structural Propagation (SP) module reinforces structural cues before prototype extraction, while a Confidence-Based Logit Refinement (CBLR) mechanism adaptively fuses base and refined prototypes. An Adversarial Dynamic Adjustment Loss (ADAL) further balances their contributions, preventing premature dominance and stabilizing optimization. Extensive experiments on S3DIS and ScanNet show that SP-DRNet consistently outperforms state-of-the-art methods. Source code is available at https://github.com/vstar37/spdrnet.

少镜头三维点云语义分割（FS-PCS）的目的是在有限的注释示例中识别新的类。现有方法面临两个关键挑战。大型场景的滑动窗口预处理通常会截断对象实例并引起结构偏差，而掩膜平均池化原型则难以捕获这些碎片化结构。此外，许多方法优先考虑查询空间精炼原型，而不是支持派生的基本原型，从而导致从原始支持结构派生的稳定的、有区别的线索的潜在妥协。为了解决这些问题，我们提出了结构传播双原型细化网络（SP-DRNet）。轻量级结构传播（SP）模块在原型提取之前加强结构线索，而基于置信度的Logit细化（CBLR）机制自适应融合基本原型和精炼原型。对抗动态调整损失（ADAL）进一步平衡他们的贡献，防止过早的优势和稳定的优化。在S3DIS和ScanNet上进行的大量实验表明，SP-DRNet始终优于最先进的方法。源代码可从https://github.com/vstar37/spdrnet获得。

{"title":"Structural propagation dual-prototype refinement network for few-shot 3D point cloud segmentation","authors":"Qingzheng Wang , Jiazhi Xie , Ning Li, Jingjun Bi, Zengwei Mai, Wenhui Liu, Xingqin Wang","doi":"10.1016/j.ins.2026.123076","DOIUrl":"10.1016/j.ins.2026.123076","url":null,"abstract":"<div><div>Few-shot 3D point cloud semantic segmentation (FS-PCS) aims to recognize novel classes with only limited annotated examples. Existing methods face two key challenges. Sliding-window preprocessing of large scenes often truncates object instances and induces structural bias, while masked average pooling prototypes struggle to capture these fragmented structures. Furthermore, many methods prioritize query-space refined prototypes over support-derived base prototypes, leading to a potential compromise of stable, discriminative cues derived from the original support structure. To address these issues, we propose the Structural Propagation Dual-Prototype Refinement Network (SP-DRNet). A lightweight Structural Propagation (SP) module reinforces structural cues before prototype extraction, while a Confidence-Based Logit Refinement (CBLR) mechanism adaptively fuses base and refined prototypes. An Adversarial Dynamic Adjustment Loss (ADAL) further balances their contributions, preventing premature dominance and stabilizing optimization. Extensive experiments on S3DIS and ScanNet show that SP-DRNet consistently outperforms state-of-the-art methods. Source code is available at <span><span>https://github.com/vstar37/spdrnet</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"735 ","pages":"Article 123076"},"PeriodicalIF":6.8,"publicationDate":"2026-01-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145940328","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

Online state recognition of streaming time series based on subsequence similarity 基于子序列相似度的流时间序列在线状态识别

IF 6.8 1区计算机科学 0 COMPUTER SCIENCE, INFORMATION SYSTEMS

Information Sciences

Pub Date : 2026-01-06 DOI: 10.1016/j.ins.2026.123082

Jie Zhang, Peng Wang, Wei Wang

The widespread deployment of sensors has generated large volumes of time series data, prompting significant interest in data mining. Real-world time series often contain an unknown number of states that reflect the underlying physical processes of the observed objects. Existing state-recognition methods primarily address static time series, while time series streams have received little attention. Online state recognition is crucial due to its broad range of practical applications. However, many approaches designed for static data cannot be directly applied to data streams due to their unbounded growth. To address these limitations, we propose OSR-TS, an approach for online state recognition in time-series streams based on subsequence similarity. OSR-TS consists of two stages. In the offline initialization stage, OSR-TS extracts representative subsequences with distinct shapes from each state and stores them as clustering prototypes. In the online stage, each subsequence in the data stream is assigned to its nearest clustering prototype, and then state recognition is performed. OSR-TS incorporates self-adaptive learning mechanisms to identify previously unknown states. Furthermore, we employ three optimization strategies to improve the computational efficiency of the data stream. Extensive experiments on real-world datasets demonstrate that OSR-TS achieves superior effectiveness and efficiency compared to state-of-the-art methods.

传感器的广泛部署产生了大量的时间序列数据，引起了人们对数据挖掘的极大兴趣。现实世界的时间序列通常包含未知数量的状态，这些状态反映了观察对象的潜在物理过程。现有的状态识别方法主要针对静态时间序列，而时间序列流很少受到关注。由于其广泛的实际应用，在线状态识别至关重要。然而，由于静态数据的无限增长，许多为静态数据设计的方法不能直接应用于数据流。为了解决这些限制，我们提出了OSR-TS，一种基于子序列相似性的时间序列流在线状态识别方法。OSR-TS包括两个阶段。在离线初始化阶段，OSR-TS从每个状态中提取具有不同形状的代表性子序列，并将其存储为聚类原型。在在线阶段，将数据流中的每个子序列分配到离它最近的聚类原型，然后进行状态识别。OSR-TS采用自适应学习机制来识别先前未知的状态。此外，我们采用了三种优化策略来提高数据流的计算效率。在真实世界数据集上进行的大量实验表明，与最先进的方法相比，OSR-TS具有更高的有效性和效率。

{"title":"Online state recognition of streaming time series based on subsequence similarity","authors":"Jie Zhang, Peng Wang, Wei Wang","doi":"10.1016/j.ins.2026.123082","DOIUrl":"10.1016/j.ins.2026.123082","url":null,"abstract":"<div><div>The widespread deployment of sensors has generated large volumes of time series data, prompting significant interest in data mining. Real-world time series often contain an unknown number of states that reflect the underlying physical processes of the observed objects. Existing state-recognition methods primarily address static time series, while time series streams have received little attention. Online state recognition is crucial due to its broad range of practical applications. However, many approaches designed for static data cannot be directly applied to data streams due to their unbounded growth. To address these limitations, we propose OSR-TS, an approach for online state recognition in time-series streams based on subsequence similarity. OSR-TS consists of two stages. In the offline initialization stage, OSR-TS extracts representative subsequences with distinct shapes from each state and stores them as clustering prototypes. In the online stage, each subsequence in the data stream is assigned to its nearest clustering prototype, and then state recognition is performed. OSR-TS incorporates self-adaptive learning mechanisms to identify previously unknown states. Furthermore, we employ three optimization strategies to improve the computational efficiency of the data stream. Extensive experiments on real-world datasets demonstrate that OSR-TS achieves superior effectiveness and efficiency compared to state-of-the-art methods.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"735 ","pages":"Article 123082"},"PeriodicalIF":6.8,"publicationDate":"2026-01-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145940388","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

Toward imperceptible 3D adversarial point clouds via gradient-guided optimization 通过梯度引导优化实现难以察觉的3D对抗点云

IF 6.8 1区计算机科学 0 COMPUTER SCIENCE, INFORMATION SYSTEMS

Information Sciences

Pub Date : 2026-01-06 DOI: 10.1016/j.ins.2025.123068

Xiaobin Wu , Jiangnan Zheng , Huan Luo

Adversarial attacks are an effective method for revealing the vulnerabilities of 3D point cloud classification models and promoting the development of more robust architectures. However, existing gradient-based attack methods often exhibit unstable update directions and slow convergence, primarily arising from the complex loss landscape associated with the unordered, sparse, and irregular structure of point cloud data, and this instability in turn compromises the imperceptibility of the generated adversarial examples. To address these challenges, a novel two-stage gradient optimization framework is proposed to generate adversarial point clouds with improved imperceptibility and optimization efficiency. Specifically, we introduce a Hybrid Gradient Descent (HGD) strategy that applies geometric transformations to create augmented samples around each adversarial point cloud. By aggregating the gradients from augmented samples, HGD effectively smooths local noise and guides the optimization toward more stable descent directions. Extending the HGD strategy, we design a Prospective Gradient Correction (PGC) mechanism that constructs forward-looking perturbation points along the current update trajectory and fuses the corresponding gradients with the original direction. PGC enables dynamic refinement of the update path, mitigating local gradient bias and aligning the update direction more accurately with the decision boundary. Experimental results show that the full HGD&PGC framework reduces required iterations by 48.5% and achieves a reduction of 18.7% in the average perturbation magnitude, all without compromising attack success rates. Furthermore, this optimization framework has the potential to be extended to other 3D data types, and also to inform defense strategies that mitigate such attacks by complicating the loss landscape.

对抗性攻击是揭示三维点云分类模型漏洞和促进更健壮架构发展的有效方法。然而，现有的基于梯度的攻击方法往往表现出不稳定的更新方向和缓慢的收敛，这主要是由于与点云数据的无序、稀疏和不规则结构相关的复杂损失情况，而这种不稳定性反过来又损害了生成的对抗性示例的不可感知性。为了解决这些问题，提出了一种新的两阶段梯度优化框架，以提高不可感知性和优化效率来生成对抗性点云。具体来说，我们引入了一种混合梯度下降（HGD）策略，该策略应用几何变换在每个对抗点云周围创建增强样本。通过对增强样本的梯度进行聚合，HGD有效地平滑了局部噪声，并将优化导向更稳定的下降方向。在HGD策略的基础上，我们设计了一种前瞻性梯度校正（PGC）机制，该机制沿着当前更新轨迹构建前瞻性扰动点，并将相应的梯度与原始方向融合。PGC支持更新路径的动态细化，减轻局部梯度偏差，并使更新方向更准确地与决策边界对齐。实验结果表明，完整的HGD&；PGC框架将所需的迭代次数减少了48.5%，平均扰动幅度减少了18.7%，而且攻击成功率不受影响。此外，该优化框架具有扩展到其他3D数据类型的潜力，并且还可以通过使损失情况复杂化来告知防御策略，以减轻此类攻击。

{"title":"Toward imperceptible 3D adversarial point clouds via gradient-guided optimization","authors":"Xiaobin Wu , Jiangnan Zheng , Huan Luo","doi":"10.1016/j.ins.2025.123068","DOIUrl":"10.1016/j.ins.2025.123068","url":null,"abstract":"<div><div>Adversarial attacks are an effective method for revealing the vulnerabilities of 3D point cloud classification models and promoting the development of more robust architectures. However, existing gradient-based attack methods often exhibit unstable update directions and slow convergence, primarily arising from the complex loss landscape associated with the unordered, sparse, and irregular structure of point cloud data, and this instability in turn compromises the imperceptibility of the generated adversarial examples. To address these challenges, a novel two-stage gradient optimization framework is proposed to generate adversarial point clouds with improved imperceptibility and optimization efficiency. Specifically, we introduce a Hybrid Gradient Descent (HGD) strategy that applies geometric transformations to create augmented samples around each adversarial point cloud. By aggregating the gradients from augmented samples, HGD effectively smooths local noise and guides the optimization toward more stable descent directions. Extending the HGD strategy, we design a Prospective Gradient Correction (PGC) mechanism that constructs forward-looking perturbation points along the current update trajectory and fuses the corresponding gradients with the original direction. PGC enables dynamic refinement of the update path, mitigating local gradient bias and aligning the update direction more accurately with the decision boundary. Experimental results show that the full HGD&PGC framework reduces required iterations by 48.5% and achieves a reduction of 18.7% in the average perturbation magnitude, all without compromising attack success rates. Furthermore, this optimization framework has the potential to be extended to other 3D data types, and also to inform defense strategies that mitigate such attacks by complicating the loss landscape.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"735 ","pages":"Article 123068"},"PeriodicalIF":6.8,"publicationDate":"2026-01-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145940391","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

Prediction of social influence in higher-order networks 高阶网络中社会影响的预测

IF 6.8 1区计算机科学 0 COMPUTER SCIENCE, INFORMATION SYSTEMS

Information Sciences

Pub Date : 2026-01-05 DOI: 10.1016/j.ins.2026.123075

Hao Peng , Rui Zhang , Bo Zhang , Cheng Qian , Ming Zhong , Shenghong Li , Jianmin Han , Dandan Zhao , Wei Wang

Social influence plays a pivotal role in understanding collective decision-making, as individuals readily observe and are influenced by their friends’ behaviours. Accurately predicting social influence at the individual level is essential for various applications, including political communication and marketing. Graph Neural Networks (GNNs) have been widely applied for modelling social influence. However, existing work often overlooks higher-order interactions and hierarchical structures. It also struggles to effectively handle the multi-type nature of social networks. In this study, we propose a framework that learns latent feature representations of users in an end-to-end manner to forecast social influence. A critical component of the framework is a novel geometric deep learning paradigm, Hyperbolic Hypergraph Convolutional Network (HHGCN), which integrates multi-space modelling and higher-order relational learning to more precisely quantify social influence. Specifically, we design a method for learning topological features and capturing group-level environmental characteristics, then transform multi-type social networks that include follow and retweet relationships into a unified hypergraph. Leveraging hyperbolic space and hyperbolic hypergraph operations, our model extracts latent predictive signals. Experiments on eight social network datasets demonstrate that HHGCN outperforms mainstream GNNs on ACC, AUC, and F1, validating the advantages of hyperbolic geometry for hierarchical structures and hypergraphs for higher-order interactions.

社会影响在理解集体决策方面起着关键作用，因为个人很容易观察并受到朋友行为的影响。准确预测个人层面的社会影响对各种应用至关重要，包括政治传播和市场营销。图神经网络（GNNs）已广泛应用于社会影响建模。然而，现有的工作往往忽略了高阶的相互作用和层次结构。它还在努力有效地处理社交网络的多类型性质。在这项研究中，我们提出了一个框架，以端到端方式学习用户的潜在特征表征来预测社会影响。该框架的一个关键组成部分是一种新的几何深度学习范式，即双曲超图卷积网络（HHGCN），它集成了多空间建模和高阶关系学习，以更精确地量化社会影响。具体而言，我们设计了一种学习拓扑特征和捕获群体级环境特征的方法，然后将包含关注和转发关系的多类型社交网络转换为统一的超图。利用双曲空间和双曲超图运算，我们的模型提取潜在的预测信号。在8个社交网络数据集上的实验表明，HHGCN在ACC、AUC和F1上优于主流gnn，验证了双曲几何在分层结构和超图在高阶交互中的优势。

{"title":"Prediction of social influence in higher-order networks","authors":"Hao Peng , Rui Zhang , Bo Zhang , Cheng Qian , Ming Zhong , Shenghong Li , Jianmin Han , Dandan Zhao , Wei Wang","doi":"10.1016/j.ins.2026.123075","DOIUrl":"10.1016/j.ins.2026.123075","url":null,"abstract":"<div><div>Social influence plays a pivotal role in understanding collective decision-making, as individuals readily observe and are influenced by their friends’ behaviours. Accurately predicting social influence at the individual level is essential for various applications, including political communication and marketing. Graph Neural Networks (GNNs) have been widely applied for modelling social influence. However, existing work often overlooks higher-order interactions and hierarchical structures. It also struggles to effectively handle the multi-type nature of social networks. In this study, we propose a framework that learns latent feature representations of users in an end-to-end manner to forecast social influence. A critical component of the framework is a novel geometric deep learning paradigm, Hyperbolic Hypergraph Convolutional Network (HHGCN), which integrates multi-space modelling and higher-order relational learning to more precisely quantify social influence. Specifically, we design a method for learning topological features and capturing group-level environmental characteristics, then transform multi-type social networks that include follow and retweet relationships into a unified hypergraph. Leveraging hyperbolic space and hyperbolic hypergraph operations, our model extracts latent predictive signals. Experiments on eight social network datasets demonstrate that HHGCN outperforms mainstream GNNs on ACC, AUC, and F1, validating the advantages of hyperbolic geometry for hierarchical structures and hypergraphs for higher-order interactions.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"735 ","pages":"Article 123075"},"PeriodicalIF":6.8,"publicationDate":"2026-01-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145940390","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

LACK: Adaptive k-means clustering with learning-augmented policy for approximate K nearest neighbor search LACK：基于学习增强策略的自适应K均值聚类近似K近邻搜索

IF 6.8 1区计算机科学 0 COMPUTER SCIENCE, INFORMATION SYSTEMS

Information Sciences

Pub Date : 2026-01-05 DOI: 10.1016/j.ins.2026.123077

Zhihao Chen, Junnuo Lin, Lingli Li, Yongnan Liu

Traditional

k

-means clustering underperforms in Approximate K Nearest Neighbor Search (AKNNS) under skewed workloads with uneven access to hot and cold data. To address this, we propose LACK (Learning-Augmented Adaptive

k

-means Clustering), a framework using learning-augmented policies to optimize clustering and query performance adaptively. LACK leverages the transformation framework (TCF), converting the Adaptive

k

-means Clustering (AKC) problem into standard

k

-means via controllable data replication. It ensures consistency of the optimal solution and approximation ratio before and after transformation. LACK introduces two key optimizations: (1) Implicit Weight Expansion (IWE), replacing explicit replication with dynamic weight management; (2) Jumping Window, dynamically adjusting computation step size based on data repetition to reduce redundancy. We also propose a Fast Sampling algorithm for efficient centroid approximation. Experiments on real datasets show LACK’s superiority over state-of-the-art

k

-means variants, maintaining excellent accuracy while significantly reducing time and increasing the recall rate of hot data queries by 10%. LACK bridges the gap between theoretical guarantees and practical efficiency in large-scale, skewed workloads.

传统的K -means聚类在热数据和冷数据访问不均匀的倾斜工作负载下，表现不佳。为了解决这个问题，我们提出了LACK（学习增强自适应k均值聚类），这是一个使用学习增强策略自适应优化聚类和查询性能的框架。LACK利用转换框架（TCF），通过可控数据复制将自适应k-均值聚类（AKC）问题转换为标准k-均值。它保证了变换前后最优解和近似比的一致性。LACK引入了两个关键优化：(1)隐式权重扩展（IWE），用动态权重管理取代显式复制；(2)跳跃窗口，根据数据重复动态调整计算步长，减少冗余。我们还提出了一种快速采样算法，用于有效的质心逼近。在真实数据集上的实验表明，LACK优于最先进的k-means变体，在保持优异准确性的同时显着减少了时间，并将热数据查询的召回率提高了10%。LACK在大规模、倾斜工作负载的理论保证和实际效率之间架起了桥梁。

{"title":"LACK: Adaptive k-means clustering with learning-augmented policy for approximate K nearest neighbor search","authors":"Zhihao Chen, Junnuo Lin, Lingli Li, Yongnan Liu","doi":"10.1016/j.ins.2026.123077","DOIUrl":"10.1016/j.ins.2026.123077","url":null,"abstract":"<div><div>Traditional <span><math><mi>k</mi></math></span>-means clustering underperforms in Approximate K Nearest Neighbor Search (AKNNS) under skewed workloads with uneven access to hot and cold data. To address this, we propose LACK (Learning-Augmented Adaptive <span><math><mi>k</mi></math></span>-means Clustering), a framework using learning-augmented policies to optimize clustering and query performance adaptively. LACK leverages the transformation framework (TCF), converting the Adaptive <span><math><mi>k</mi></math></span>-means Clustering (AKC) problem into standard <span><math><mi>k</mi></math></span>-means via controllable data replication. It ensures consistency of the optimal solution and approximation ratio before and after transformation. LACK introduces two key optimizations: (1) Implicit Weight Expansion (IWE), replacing explicit replication with dynamic weight management; (2) Jumping Window, dynamically adjusting computation step size based on data repetition to reduce redundancy. We also propose a Fast Sampling algorithm for efficient centroid approximation. Experiments on real datasets show LACK’s superiority over state-of-the-art <span><math><mi>k</mi></math></span>-means variants, maintaining excellent accuracy while significantly reducing time and increasing the recall rate of hot data queries by 10%. LACK bridges the gap between theoretical guarantees and practical efficiency in large-scale, skewed workloads.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"735 ","pages":"Article 123077"},"PeriodicalIF":6.8,"publicationDate":"2026-01-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145940325","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

GS-UNet: ConvNeXt-based keypoint-driven visual servoing with cross-hierarchical attention gating for high-precision robotic assembly GS-UNet：基于convnext的关键点驱动视觉伺服与交叉分层注意门控，用于高精度机器人装配

IF 6.8 1区计算机科学 0 COMPUTER SCIENCE, INFORMATION SYSTEMS

Information Sciences

Pub Date : 2026-01-05 DOI: 10.1016/j.ins.2026.123079

Fengze Xu, Liming Lao, Xuan Zheng, Yongchao Zhang, Pengzhan Chen

Achieving high-precision robotic assembly in unstructured environments remains highly challenging. To overcome the heavy manual tuning and limited robustness of conventional visual servoing, as well as the accuracy bottlenecks and pipeline redundancy of existing deep learning–based approaches, we propose Gated Servo U-Net (GS-UNet). GS-UNet adopts ConvNeXt as the encoder and introduces a ConvNeXt-homologous attention gate (CHAG) for cross-hierarchical fusion of multi-scale features and selective focus on key geometric structures. Trained end-to-end via a synthetic data pipeline based on domain randomization (DR), the network directly regresses task keypoints to drive image-based visual servoing (IBVS).

On a real robotic platform, we systematically evaluate GS-UNet in five physical scenarios, including a nominal setting, clutter and occlusion, low light, varying exposure parameters, and tight-clearance assembly under base perturbations, for a total of 130 insertion trials. Overall, the success rates across these scenarios range from 75.00% to 100%, with mean end-effector position errors of approximately 0.27–0.49 mm and mean orientation errors of approximately 0.085–0.24

^{\circ}

. These results demonstrate that the proposed system can maintain sub-millimeter assembly accuracy and strong robustness under complex and variable operating conditions in realistic unstructured environments.

在非结构化环境中实现高精度机器人装配仍然是极具挑战性的。为了克服传统视觉伺服的大量手动调谐和有限的鲁棒性，以及现有基于深度学习的方法的精度瓶颈和管道冗余，我们提出了门控伺服U-Net （GS-UNet）。GS-UNet采用ConvNeXt作为编码器，引入了一种ConvNeXt-homologous attention gate (CHAG)，用于多尺度特征的跨层次融合和对关键几何结构的选择性聚焦。该网络通过基于领域随机化（DR）的综合数据管道进行端到端训练，直接回归任务关键点来驱动基于图像的视觉伺服（IBVS）。在一个真实的机器人平台上，我们系统地评估了GS-UNet在五种物理场景下的性能，包括名义设置、杂波和遮挡、低光、不同的曝光参数以及在基础扰动下的紧密间隙组装，总共进行了130次插入试验。总的来说，这些场景的成功率从75.00%到100%不等，平均执行器位置误差约为0.27-0.49 mm，平均方向误差约为0.085-0.24°。结果表明，该系统在复杂多变的实际非结构化环境下仍能保持亚毫米级装配精度和较强的鲁棒性。

{"title":"GS-UNet: ConvNeXt-based keypoint-driven visual servoing with cross-hierarchical attention gating for high-precision robotic assembly","authors":"Fengze Xu, Liming Lao, Xuan Zheng, Yongchao Zhang, Pengzhan Chen","doi":"10.1016/j.ins.2026.123079","DOIUrl":"10.1016/j.ins.2026.123079","url":null,"abstract":"<div><div>Achieving high-precision robotic assembly in unstructured environments remains highly challenging. To overcome the heavy manual tuning and limited robustness of conventional visual servoing, as well as the accuracy bottlenecks and pipeline redundancy of existing deep learning–based approaches, we propose Gated Servo U-Net (GS-UNet). GS-UNet adopts ConvNeXt as the encoder and introduces a ConvNeXt-homologous attention gate (CHAG) for cross-hierarchical fusion of multi-scale features and selective focus on key geometric structures. Trained end-to-end via a synthetic data pipeline based on domain randomization (DR), the network directly regresses task keypoints to drive image-based visual servoing (IBVS).</div><div>On a real robotic platform, we systematically evaluate GS-UNet in five physical scenarios, including a nominal setting, clutter and occlusion, low light, varying exposure parameters, and tight-clearance assembly under base perturbations, for a total of 130 insertion trials. Overall, the success rates across these scenarios range from 75.00% to 100%, with mean end-effector position errors of approximately 0.27–0.49 mm and mean orientation errors of approximately 0.085–0.24<span><math><msup><mspace></mspace><mrow><mo>∘</mo></mrow></msup></math></span>. These results demonstrate that the proposed system can maintain sub-millimeter assembly accuracy and strong robustness under complex and variable operating conditions in realistic unstructured environments.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"735 ","pages":"Article 123079"},"PeriodicalIF":6.8,"publicationDate":"2026-01-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145940329","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

Resilient tracking control of UAV with event-triggered communication against stochastic DoS attacks and faults 针对随机DoS攻击和故障的事件触发通信无人机弹性跟踪控制

IF 6.8 1区计算机科学 0 COMPUTER SCIENCE, INFORMATION SYSTEMS

Information Sciences

Pub Date : 2026-01-03 DOI: 10.1016/j.ins.2026.123074

Minrui Fu, Ziquan Yu

This article investigates a resilient control strategy for an unmanned aerial vehicle (UAV) under an event-triggered communication (ETC) mechanism to simultaneously resist stochastic denial-of-service (DoS) attacks and faults. Firstly, the network health state during stochastic DoS attacks is modeled as a Markov process by using random variables to represent both normal and interrupted periods. Secondly, an observer incorporating a radial basis function neural network (RBFNN) is developed to estimate unknown components caused by faults and uncertain dynamics. Then, auxiliary control laws are proposed for fault compensation. Furthermore, a resilient controller is designed to enhance system resilience, along with a dynamic ETC mechanism that adaptively adjusts the trigger frequency. Moreover, a stochastic hybrid system model of the UAV is constructed to better incorporate the continuous states, jump states and the random inputs. Within this hybrid system framework, it is proved that the tracking error remains Lagrange stable in probability and Lyapunov stable in probability. Finally, the feasibility and efficiency of the proposed scheme are validated through a hardware-in-the-loop experiment using the open-source flight autopilot Pixhawk

^{®}

6C.

本文研究了一种事件触发通信（ETC）机制下的无人机弹性控制策略，以同时抵抗随机拒绝服务（DoS）攻击和故障。首先，将随机DoS攻击时的网络健康状态建模为马尔可夫过程，使用随机变量表示正常和中断时间段；其次，提出了一种基于径向基函数神经网络（RBFNN）的观测器，用于估计由故障和不确定动态引起的未知分量；然后，提出了辅助控制律进行故障补偿。此外，设计了弹性控制器来增强系统的弹性，以及自适应调整触发频率的动态ETC机制。在此基础上，构建了无人机的随机混合系统模型，以更好地融合连续状态、跳跃状态和随机输入。在此混合系统框架下，证明了跟踪误差在概率上保持拉格朗日稳定，在概率上保持李雅普诺夫稳定。最后，利用开源飞行自动驾驶仪Pixhawk®6C进行了硬件在环实验，验证了该方案的可行性和效率。

{"title":"Resilient tracking control of UAV with event-triggered communication against stochastic DoS attacks and faults","authors":"Minrui Fu, Ziquan Yu","doi":"10.1016/j.ins.2026.123074","DOIUrl":"10.1016/j.ins.2026.123074","url":null,"abstract":"<div><div>This article investigates a resilient control strategy for an unmanned aerial vehicle (UAV) under an event-triggered communication (ETC) mechanism to simultaneously resist stochastic denial-of-service (DoS) attacks and faults. Firstly, the network health state during stochastic DoS attacks is modeled as a Markov process by using random variables to represent both normal and interrupted periods. Secondly, an observer incorporating a radial basis function neural network (RBFNN) is developed to estimate unknown components caused by faults and uncertain dynamics. Then, auxiliary control laws are proposed for fault compensation. Furthermore, a resilient controller is designed to enhance system resilience, along with a dynamic ETC mechanism that adaptively adjusts the trigger frequency. Moreover, a stochastic hybrid system model of the UAV is constructed to better incorporate the continuous states, jump states and the random inputs. Within this hybrid system framework, it is proved that the tracking error remains Lagrange stable in probability and Lyapunov stable in probability. Finally, the feasibility and efficiency of the proposed scheme are validated through a hardware-in-the-loop experiment using the open-source flight autopilot Pixhawk<span><math><msup><mspace></mspace><mi>®</mi></msup></math></span> 6C.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"735 ","pages":"Article 123074"},"PeriodicalIF":6.8,"publicationDate":"2026-01-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145940334","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