Neurocomputing最新文献_第8页

Eliminating Information Cocoons: Product Diversification Recommendation Based on User Clustering and Hybrid Reranking 消除信息茧：基于用户聚类和混合重排序的产品多样化推荐

IF 6.5 2区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neurocomputing

Pub Date : 2026-01-02 DOI: 10.1016/j.neucom.2025.132595

Jiaxin Li , Jinyun Fang

Recommendation systems, while alleviating information overload, often over-specialize and trap users in “information cocoons.” To address this, we propose a novel two-stage diversified recommendation framework that strategically separates accuracy optimization from diversity enhancement. In the first stage, we construct a clustering-based local-hybrid model (RHM). It fuses a global model, built on a genre-augmented rating matrix filled via the Weighted Slope One (WSO) algorithm, with local models derived from user clusters identified via Latent Dirichlet Allocation (LDA). This stage establishes a robust foundation of recommendation accuracy. In the second stage, we introduce a hybrid reranking strategy with an adaptive switching mechanism. For each user, it dynamically chooses between a threshold reranking method (which penalizes both item and genre popularity to boost genre coverage) and a greedy reranking method (which incorporates a binomial diversity framework to jointly optimize relevance and genre coverage). This stage is dedicated to diversity enhancement with minimal accuracy loss. Guided by the philosophy that breaking deep information filters requires dedicated, sequential optimization of accuracy and diversity, our framework offers a clear pathway toward more open recommendations. Experiments on a movie dataset show that RHM improves recommendation accuracy (nDCG) by 32.4 % over a standard UserCF baseline. The full diversified model (DRHM) further enhances genre coverage (GC) by 13.7 % and overall coverage (COV) by 56.1 %, while retaining 96.7 % of RHM's accuracy. The proposed framework effectively balances the accuracy-diversity trade-off, offering a practical pathway toward more open and equitable recommendation ecosystems.

推荐系统虽然减轻了信息过载，但往往过于专业化，将用户困在“信息茧”中。为了解决这个问题，我们提出了一种新的两阶段多样化推荐框架，该框架将准确性优化与多样性增强战略性地分离开来。在第一阶段，我们构建了一个基于聚类的局部混合模型。它融合了一个建立在通过加权斜率1 （WSO）算法填充的类型增强评级矩阵上的全局模型，以及通过潜在狄利克雷分配（LDA）识别的用户集群派生的局部模型。这一阶段为推荐的准确性奠定了坚实的基础。在第二阶段，我们引入了一种具有自适应切换机制的混合重排序策略。对于每个用户，它会动态选择阈值重新排名方法（惩罚道具和类型的流行度以提高类型的覆盖率）和贪婪重新排名方法（结合二项多样性框架以共同优化相关性和类型覆盖率）。这一阶段致力于以最小的精度损失来增强分集。在打破深层信息过滤器需要专门的、顺序的准确性和多样性优化这一理念的指导下，我们的框架为更开放的推荐提供了一条清晰的途径。在电影数据集上的实验表明，与标准UserCF基线相比，RHM将推荐准确率（nDCG）提高了32.4% %。完全多样化模型（DRHM）进一步提高了类型覆盖率（GC） 13.7% %，总体覆盖率（COV） 56.1% %，同时保持了RHM的96.7 %的准确率。该框架有效地平衡了准确性和多样性之间的权衡，为构建更加开放和公平的推荐生态系统提供了一条切实可行的途径。

{"title":"Eliminating Information Cocoons: Product Diversification Recommendation Based on User Clustering and Hybrid Reranking","authors":"Jiaxin Li , Jinyun Fang","doi":"10.1016/j.neucom.2025.132595","DOIUrl":"10.1016/j.neucom.2025.132595","url":null,"abstract":"<div><div>Recommendation systems, while alleviating information overload, often over-specialize and trap users in “information cocoons.” To address this, we propose a novel two-stage diversified recommendation framework that strategically separates accuracy optimization from diversity enhancement. In the first stage, we construct a clustering-based local-hybrid model (RHM). It fuses a global model, built on a genre-augmented rating matrix filled via the Weighted Slope One (WSO) algorithm, with local models derived from user clusters identified via Latent Dirichlet Allocation (LDA). This stage establishes a robust foundation of recommendation accuracy. In the second stage, we introduce a hybrid reranking strategy with an adaptive switching mechanism. For each user, it dynamically chooses between a threshold reranking method (which penalizes both item and genre popularity to boost genre coverage) and a greedy reranking method (which incorporates a binomial diversity framework to jointly optimize relevance and genre coverage). This stage is dedicated to diversity enhancement with minimal accuracy loss. Guided by the philosophy that breaking deep information filters requires dedicated, sequential optimization of accuracy and diversity, our framework offers a clear pathway toward more open recommendations. Experiments on a movie dataset show that RHM improves recommendation accuracy (nDCG) by 32.4 % over a standard UserCF baseline. The full diversified model (DRHM) further enhances genre coverage (GC) by 13.7 % and overall coverage (COV) by 56.1 %, while retaining 96.7 % of RHM's accuracy. The proposed framework effectively balances the accuracy-diversity trade-off, offering a practical pathway toward more open and equitable recommendation ecosystems.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"670 ","pages":"Article 132595"},"PeriodicalIF":6.5,"publicationDate":"2026-01-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145886509","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"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

DAPU: Distribution-aware patch upsampling for point cloud-based 3D object detection DAPU：基于点云的3D物体检测的分布感知补丁采样

IF 6.5 2区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neurocomputing

Pub Date : 2026-01-02 DOI: 10.1016/j.neucom.2025.132579

Yinghao Hu, Yan Wu, Yujian Mo, Jijun Wang, Yuwei Zhang

The sparsity and quality of point clouds significantly constrain the development of LiDAR-based 3D object detectors. Previous approaches supplemented point clouds through depth completion or upsampling. However, the former suffers from inconsistencies caused by differences in multimodal data, resulting in uneven point cloud quality. Meanwhile, previous upsampling methods convert point clouds into range images which results in a loss of point accuracy. In this paper, we present DAPU, a novel real-time point cloud upsampling method designed to address these challenges. This method consists of three key components: (1) the Ground Points Recognizer (GPR), which analyzes the height difference distribution between coplanar and non-coplanar points within patches to identify ground points, establishes a sparse-to-dense index matrix for fast large-scale point cloud queries; (2) the Distribution-Aware Patch KNN (DAPKNN), which dynamically adjusts the sampling radius threshold based on distribution to reduce computation and ensure sufficient neighbor sampling for distant points; (3) the Neighbors Upsampling, which linearly upsamples between each pair of neighbors to preserve all point features. KITTI experiments show gains of up to +1.2% AP 3D and +1.4% AP BEV. Additional evaluations on mini-nuScenes and Waymo further demonstrate consistent improvements across Vehicle, Pedestrian, and Cyclist detection, confirming DAPU’s robustness under diverse LiDAR settings and real-time suitability.

点云的稀疏性和质量严重制约了基于激光雷达的三维目标探测器的发展。以前的方法通过深度补全或上采样来补充点云。然而，前者由于多模态数据的差异而存在不一致性，导致点云质量参差不齐。同时，以往的上采样方法将点云转换为距离图像，导致点精度下降。在本文中，我们提出了DAPU，一种新的实时点云上采样方法，旨在解决这些挑战。该方法由三个关键部分组成：(1)地点识别器（GPR）通过分析斑块内共面点与非共面点的高差分布来识别地点，建立从稀疏到密集的索引矩阵，用于快速大规模点云查询；(2)分布感知Patch KNN (DAPKNN)，该算法基于分布动态调整采样半径阈值，以减少计算量并保证对距离较远的点进行足够的邻居采样；(3)邻居上采样，在每对邻居之间进行线性上采样，以保留所有的点特征。KITTI实验显示，AP 3D和AP BEV的增益分别为+1.2%和+1.4%。对mini-nuScenes和Waymo的额外评估进一步证明了在车辆、行人和骑自行车者检测方面的一致性改进，证实了DAPU在不同激光雷达设置下的稳健性和实时适用性。

{"title":"DAPU: Distribution-aware patch upsampling for point cloud-based 3D object detection","authors":"Yinghao Hu, Yan Wu, Yujian Mo, Jijun Wang, Yuwei Zhang","doi":"10.1016/j.neucom.2025.132579","DOIUrl":"10.1016/j.neucom.2025.132579","url":null,"abstract":"<div><div>The sparsity and quality of point clouds significantly constrain the development of LiDAR-based 3D object detectors. Previous approaches supplemented point clouds through depth completion or upsampling. However, the former suffers from inconsistencies caused by differences in multimodal data, resulting in uneven point cloud quality. Meanwhile, previous upsampling methods convert point clouds into range images which results in a loss of point accuracy. In this paper, we present DAPU, a novel real-time point cloud upsampling method designed to address these challenges. This method consists of three key components: (1) the Ground Points Recognizer (GPR), which analyzes the height difference distribution between coplanar and non-coplanar points within patches to identify ground points, establishes a sparse-to-dense index matrix for fast large-scale point cloud queries; (2) the Distribution-Aware Patch KNN (DAPKNN), which dynamically adjusts the sampling radius threshold based on distribution to reduce computation and ensure sufficient neighbor sampling for distant points; (3) the Neighbors Upsampling, which linearly upsamples between each pair of neighbors to preserve all point features. KITTI experiments show gains of up to +1.2% AP 3D and +1.4% AP BEV. Additional evaluations on mini-nuScenes and Waymo further demonstrate consistent improvements across Vehicle, Pedestrian, and Cyclist detection, confirming DAPU’s robustness under diverse LiDAR settings and real-time suitability.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"670 ","pages":"Article 132579"},"PeriodicalIF":6.5,"publicationDate":"2026-01-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145928170","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"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

Prescribed performance safe optimal tracking control for nonlinear systems subject to dynamic obstacles via NN-based adaptive dynamic programming 基于神经网络自适应动态规划的动态障碍非线性系统规定性能安全最优跟踪控制

IF 6.5 2区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neurocomputing

Pub Date : 2026-01-02 DOI: 10.1016/j.neucom.2025.132600

Yiyu Feng, Weihao Pan, Xianfu Zhang

This paper investigates the prescribed performance safe optimal tracking control problem for uncertain systems with both unknown nonlinearities and unknown disturbances in a multi-dynamic obstacle environment. Different from existing studies, this paper proposes a feedforward-feedback optimal control strategy by combining neural network-based adaptive dynamic programming with backstepping techniques, thereby simultaneously addressing the multiple challenges of prescribed performance tracking, obstacle avoidance, and cost optimization. Specifically, based on a unified integral-multiplicative tangent barrier Lyapunov function, a feedforward controller is designed to address prescribed performance constraint for tracking error and obstacle avoidance requirements. Subsequently, by integrating a neural network approximation of the solution to the Hamilton-Jacobi-Bellman equation with a parameter adaptive law, an adaptive feedback optimal controller is obtained. Additionally, through solving the differential game problem of the affine system, the unknown disturbances are incorporated into the design framework. Rigorous stability analysis demonstrates that all signals in the closed-loop system are uniformly ultimately bounded and that the prescribed performance optimal tracking with obstacle avoidance is achieved. Finally, simulation results validate the effectiveness of the proposed control strategy.

研究了多动态障碍环境下具有未知非线性和未知扰动的不确定系统的规定性能安全最优跟踪控制问题。与已有研究不同，本文将基于神经网络的自适应动态规划与反演技术相结合，提出了一种前馈-反馈最优控制策略，从而同时解决了规定性能跟踪、避障和成本优化的多重挑战。具体而言，基于统一的积分乘切线障碍Lyapunov函数，设计了前馈控制器，以解决跟踪误差和避障要求的规定性能约束。然后，将Hamilton-Jacobi-Bellman方程解的神经网络逼近与参数自适应律进行积分，得到自适应反馈最优控制器。此外，通过求解仿射系统的微分博弈问题，将未知干扰纳入设计框架。严格的稳定性分析表明，闭环系统的所有信号最终都是一致有界的，并实现了规定性能的最优避障跟踪。最后，仿真结果验证了所提控制策略的有效性。

{"title":"Prescribed performance safe optimal tracking control for nonlinear systems subject to dynamic obstacles via NN-based adaptive dynamic programming","authors":"Yiyu Feng, Weihao Pan, Xianfu Zhang","doi":"10.1016/j.neucom.2025.132600","DOIUrl":"10.1016/j.neucom.2025.132600","url":null,"abstract":"<div><div>This paper investigates the prescribed performance safe optimal tracking control problem for uncertain systems with both unknown nonlinearities and unknown disturbances in a multi-dynamic obstacle environment. Different from existing studies, this paper proposes a feedforward-feedback optimal control strategy by combining neural network-based adaptive dynamic programming with backstepping techniques, thereby simultaneously addressing the multiple challenges of prescribed performance tracking, obstacle avoidance, and cost optimization. Specifically, based on a unified integral-multiplicative tangent barrier Lyapunov function, a feedforward controller is designed to address prescribed performance constraint for tracking error and obstacle avoidance requirements. Subsequently, by integrating a neural network approximation of the solution to the Hamilton-Jacobi-Bellman equation with a parameter adaptive law, an adaptive feedback optimal controller is obtained. Additionally, through solving the differential game problem of the affine system, the unknown disturbances are incorporated into the design framework. Rigorous stability analysis demonstrates that all signals in the closed-loop system are uniformly ultimately bounded and that the prescribed performance optimal tracking with obstacle avoidance is achieved. Finally, simulation results validate the effectiveness of the proposed control strategy.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"670 ","pages":"Article 132600"},"PeriodicalIF":6.5,"publicationDate":"2026-01-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145928296","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"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

RankSAM: Lightweight adapters and prompt generation in zero-shot semantic segmentation RankSAM：零采样语义分割中的轻量级适配器和提示符生成

IF 6.5 2区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neurocomputing

Pub Date : 2026-01-02 DOI: 10.1016/j.neucom.2025.132594

Yue Zhuo , Zhaocheng Xu , Di Zhou , Pengpeng Xu , Yan Tian

Zero-shot segmentation plays a crucial role in neurocomputing, such as embodied intelligence systems and autonomous driving technologies. However, current approaches struggle to preserve the intrinsic generalization ability of SAM as input quality declines. In addition, prompt generation still faces a challenge in balancing effectiveness and efficiency. Motivated by low-rank adaptation (LoRA), we design RankSAM, which integrates slim, adaptable modules into the middle layers of the frozen SAM framework. These modules dynamically fine-tune the operational rank of their weight matrices in response to input data, leveraging a trainable gating mechanism to selectively activate specific (rank-1) matrix components as needed. In addition, a learnable prompt predictor is designed to learn and generate prompt confidence maps and point prompts, and any remaining prompts that would produce the same mask are filtered out to enhance efficiency in prompt generation. The experimental results on multiple datasets indicate that our approach improves the mean intersection over union (mIoU) by a margin of 2.5%–2.8% compared to the prevailing approaches. Project page: https://messeyamumu.github.io/RankSAM.

Zero-shot segmentation在具身智能系统和自动驾驶技术等神经计算领域发挥着至关重要的作用。然而，随着输入质量的下降，目前的方法很难保持SAM固有的泛化能力。此外，提示生成仍然面临着平衡有效性和效率的挑战。在低阶自适应（LoRA）的激励下，我们设计了RankSAM，它将精简、可适应的模块集成到冻结SAM框架的中间层中。这些模块根据输入数据动态微调其权重矩阵的操作等级，利用可训练的门控机制根据需要选择性地激活特定的（rank-1）矩阵组件。此外，设计了一个可学习的提示预测器来学习和生成提示置信度图和点提示，并过滤掉任何可能产生相同掩码的提示，以提高提示生成的效率。在多个数据集上的实验结果表明，与现有方法相比，我们的方法将平均交联（mIoU）提高了2.5%-2.8%。项目页面：https://messeyamumu.github.io/RankSAM。

{"title":"RankSAM: Lightweight adapters and prompt generation in zero-shot semantic segmentation","authors":"Yue Zhuo , Zhaocheng Xu , Di Zhou , Pengpeng Xu , Yan Tian","doi":"10.1016/j.neucom.2025.132594","DOIUrl":"10.1016/j.neucom.2025.132594","url":null,"abstract":"<div><div>Zero-shot segmentation plays a crucial role in neurocomputing, such as embodied intelligence systems and autonomous driving technologies. However, current approaches struggle to preserve the intrinsic generalization ability of SAM as input quality declines. In addition, prompt generation still faces a challenge in balancing effectiveness and efficiency. Motivated by low-rank adaptation (LoRA), we design RankSAM, which integrates slim, adaptable modules into the middle layers of the frozen SAM framework. These modules dynamically fine-tune the operational rank of their weight matrices in response to input data, leveraging a trainable gating mechanism to selectively activate specific (rank-1) matrix components as needed. In addition, a learnable prompt predictor is designed to learn and generate prompt confidence maps and point prompts, and any remaining prompts that would produce the same mask are filtered out to enhance efficiency in prompt generation. The experimental results on multiple datasets indicate that our approach improves the mean intersection over union (mIoU) by a margin of 2.5%–2.8% compared to the prevailing approaches. Project page: <span><span>https://messeyamumu.github.io/RankSAM</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"670 ","pages":"Article 132594"},"PeriodicalIF":6.5,"publicationDate":"2026-01-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145928301","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"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

DenseBAM-GI: Attention augmented DenseNet with momentum aided GRU for HMER DenseBAM-GI：注意力增强DenseNet与动量辅助GRU用于HMER

IF 6.5 2区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neurocomputing

Pub Date : 2026-01-02 DOI: 10.1016/j.neucom.2025.132572

Aniket Pal, Krishna Pratap Singh

The task of recognizing Handwritten Mathematical Expressions (HMER) is crucial in the fields of digital education and scholarly research. However, it is difficult to accurately determine the length and complex spatial relationships among symbols in handwritten mathematical expressions. In this study, we present a novel encoder–decoder architecture (DenseBAM-GI) for HMER, where the encoder has a Bottleneck Attention Module (BAM) to improve feature representation and the decoder has a Gated Input-GRU (GI-GRU) unit with an extra gate to make decoding long and complex expressions easier. The proposed model is an efficient and lightweight architecture with performance equivalent to state-of-the-art models in terms of Expression Recognition Rate (exprate). It also performs better in terms of top 1, 2, and 3 error accuracy (

\leq 1 (%)

,

\leq 2 (%)

and

\leq 3 (%)

) across the CROHME 2014, 2016, and 2019 datasets. DenseBAM-GI achieves the best exprate among all models on the CROHME 2019 dataset. Importantly, these successes are accomplished with a drop in the complexity of the calculation and a reduction in the need for GPU memory.

手写数学表达式（HMER）的识别任务在数字教育和学术研究领域至关重要。然而，手写体数学表达式中符号之间的长度和复杂的空间关系难以准确确定。在本研究中，我们为HMER提出了一种新颖的编码器-解码器架构（DenseBAM-GI），其中编码器具有瓶颈注意模块（BAM）以改善特征表示，解码器具有带额外门控的输入- gru （GI-GRU）单元，使解码长而复杂的表达式更容易。所提出的模型是一种高效且轻量级的体系结构，其性能在表达式识别率（expate）方面与最先进的模型相当。在CROHME 2014、2016和2019数据集上，它在前1、2和3的误差精度（≤1（%）、≤2（%）和≤3(%)）方面也表现得更好。DenseBAM-GI在CROHME 2019数据集上的所有模型中获得了最好的计算结果。重要的是，这些成功是在计算复杂性下降和GPU内存需求减少的情况下实现的。

{"title":"DenseBAM-GI: Attention augmented DenseNet with momentum aided GRU for HMER","authors":"Aniket Pal, Krishna Pratap Singh","doi":"10.1016/j.neucom.2025.132572","DOIUrl":"10.1016/j.neucom.2025.132572","url":null,"abstract":"<div><div>The task of recognizing Handwritten Mathematical Expressions (HMER) is crucial in the fields of digital education and scholarly research. However, it is difficult to accurately determine the length and complex spatial relationships among symbols in handwritten mathematical expressions. In this study, we present a novel encoder–decoder architecture (DenseBAM-GI) for HMER, where the encoder has a Bottleneck Attention Module (BAM) to improve feature representation and the decoder has a Gated Input-GRU (GI-GRU) unit with an extra gate to make decoding long and complex expressions easier. The proposed model is an efficient and lightweight architecture with performance equivalent to state-of-the-art models in terms of Expression Recognition Rate (exprate). It also performs better in terms of top 1, 2, and 3 error accuracy (<span><math><mo>≤</mo><mn>1</mn><mo>(</mo><mi>%</mi><mo>)</mo></math></span>, <span><math><mo>≤</mo><mn>2</mn><mo>(</mo><mi>%</mi><mo>)</mo></math></span> and <span><math><mo>≤</mo><mn>3</mn><mo>(</mo><mi>%</mi><mo>)</mo></math></span>) across the CROHME 2014, 2016, and 2019 datasets. DenseBAM-GI achieves the best exprate among all models on the CROHME 2019 dataset. Importantly, these successes are accomplished with a drop in the complexity of the calculation and a reduction in the need for GPU memory.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"670 ","pages":"Article 132572"},"PeriodicalIF":6.5,"publicationDate":"2026-01-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145928297","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"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

Text-to-video person re-identification benchmark: Dataset and dual-modal contextual alignment 文本到视频的人物再识别基准：数据集和双模态上下文对齐

IF 6.5 2区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neurocomputing

Pub Date : 2026-01-02 DOI: 10.1016/j.neucom.2025.132596

Jiajun Su , Simin Zhan , Pudu Liu , Jianqing Zhu , Huanqiang Zeng

Text-to-video person re-identification aims to identify individuals in video sequences based on textual descriptions, but it faces challenges such as insufficient large-scale annotated datasets and inadequate cross-modal alignment mechanisms. As a result, current state-of-the-art (SOTA) methods are typically trained and evaluated on small-scale benchmarks and rely on coarse cross-modal alignment strategies, which limit their ability to generalize and fully exploit temporal information. Current methods struggle to bridge the semantic gap between static text and dynamic video content, particularly in capturing temporal dynamics and fine-grained spatial–temporal correspondences. These issues constitute key technological limitations of existing SOTA approaches and directly motivate the development of both a larger, more realistic benchmark and a more effective cross-modal alignment mechanism. To address these issues, we introduce TV-MARS, a benchmark built on the MARS dataset with 16,360 text–video pairs, enriched with natural language annotations describing motion states and environmental interactions. At approximately 4.8

\times

larger than existing text-to-video person re-identification benchmarks, TV-MARS provides a more comprehensive resource for research. Additionally, we propose a Dual-Modal Contextual Alignment (DMCA) method to bridge the modality gap between text and video sequences. DMCA employs a local contextualizer to extract fine-grained spatial features and a global integrator to synthesize temporal dynamics, adaptively fusing these features to create a unified representation that aligns static textual descriptions with dynamic video content, ensuring robust cross-modal semantic consistency. Experiments show that DMCA achieves an 8.88% improvement in Rank-1 accuracy, significantly advancing the state of the art in text-to-video person re-identification.

文本-视频人物再识别旨在基于文本描述识别视频序列中的个体，但它面临着大规模标注数据集不足和跨模态对齐机制不完善等挑战。因此，当前最先进的（SOTA）方法通常是在小规模基准上进行训练和评估的，并且依赖于粗糙的跨模态对齐策略，这限制了它们泛化和充分利用时间信息的能力。目前的方法难以弥合静态文本和动态视频内容之间的语义差距，特别是在捕获时间动态和细粒度时空对应方面。这些问题构成了现有SOTA方法的关键技术限制，并直接推动了更大、更现实的基准和更有效的跨模态对齐机制的发展。为了解决这些问题，我们引入了TV-MARS，这是一个基于MARS数据集的基准测试，该数据集包含16360个文本视频对，丰富了描述运动状态和环境交互的自然语言注释。TV-MARS比现有的文本到视频人员再识别基准大约4.8倍，为研究提供了更全面的资源。此外，我们提出了一种双模态上下文对齐（DMCA）方法来弥合文本和视频序列之间的模态差距。DMCA使用本地上下文程序提取细粒度的空间特征，使用全局积分器合成时间动态，自适应地融合这些特征以创建统一的表示，使静态文本描述与动态视频内容保持一致，确保鲁棒的跨模态语义一致性。实验表明，DMCA在Rank-1准确率上提高了8.88%，显著提高了文本到视频人物再识别的技术水平。

{"title":"Text-to-video person re-identification benchmark: Dataset and dual-modal contextual alignment","authors":"Jiajun Su , Simin Zhan , Pudu Liu , Jianqing Zhu , Huanqiang Zeng","doi":"10.1016/j.neucom.2025.132596","DOIUrl":"10.1016/j.neucom.2025.132596","url":null,"abstract":"<div><div>Text-to-video person re-identification aims to identify individuals in video sequences based on textual descriptions, but it faces challenges such as insufficient large-scale annotated datasets and inadequate cross-modal alignment mechanisms. As a result, current state-of-the-art (SOTA) methods are typically trained and evaluated on small-scale benchmarks and rely on coarse cross-modal alignment strategies, which limit their ability to generalize and fully exploit temporal information. Current methods struggle to bridge the semantic gap between static text and dynamic video content, particularly in capturing temporal dynamics and fine-grained spatial–temporal correspondences. These issues constitute key technological limitations of existing SOTA approaches and directly motivate the development of both a larger, more realistic benchmark and a more effective cross-modal alignment mechanism. To address these issues, we introduce TV-MARS, a benchmark built on the MARS dataset with 16,360 text–video pairs, enriched with natural language annotations describing motion states and environmental interactions. At approximately 4.8 <span><math><mo>×</mo></math></span> larger than existing text-to-video person re-identification benchmarks, TV-MARS provides a more comprehensive resource for research. Additionally, we propose a Dual-Modal Contextual Alignment (DMCA) method to bridge the modality gap between text and video sequences. DMCA employs a local contextualizer to extract fine-grained spatial features and a global integrator to synthesize temporal dynamics, adaptively fusing these features to create a unified representation that aligns static textual descriptions with dynamic video content, ensuring robust cross-modal semantic consistency. Experiments show that DMCA achieves an 8.88% improvement in Rank-1 accuracy, significantly advancing the state of the art in text-to-video person re-identification.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"670 ","pages":"Article 132596"},"PeriodicalIF":6.5,"publicationDate":"2026-01-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145928166","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"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

PriFLRC: A secure multi-party computation-based privacy-enhanced federated learning scheme resilient to collusion PriFLRC：一种安全的基于多方计算的隐私增强联合学习方案

IF 6.5 2区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neurocomputing

Pub Date : 2026-01-02 DOI: 10.1016/j.neucom.2025.132574

Anh-Tu Tran , The-Dung Luong , Van-Nam Huynh

By leveraging deep learning-based technologies, a range of complex problems across various domains has been effectively addressed. However, privacy concerns render traditional centralized training methods unsuitable for sensitive data-driven scenarios, such as those in healthcare and finance. Recently, federated learning has attracted significant attention for enabling participants to collaboratively train a shared model without disclosing their local data. Nonetheless, research has shown that adversaries can still exploit the shared parameters to compromise the applications. To address this issue, we propose PriFLRC, an efficient and privacy-enhanced federated learning scheme. Unlike existing solutions, PriFLRC can prevent private data leaks even in the presence of colluding entities, without the need for a fully trusted party. Experimental evaluations conducted on multiple datasets, including MNIST, CSIC 2010, SMS Spam and CIFAR-10, demonstrate that PriFLRC excels in both accuracy and efficiency. These results lay the groundwork for implementing privacy-preserving machine learning applications in real-world scenarios.

通过利用基于深度学习的技术，各种领域的一系列复杂问题得到了有效解决。然而，出于隐私方面的考虑，传统的集中式培训方法不适合敏感的数据驱动场景，例如医疗保健和金融领域。最近，联邦学习引起了人们的极大关注，因为它使参与者能够在不泄露其本地数据的情况下协作训练共享模型。尽管如此，研究表明，攻击者仍然可以利用共享参数来破坏应用程序。为了解决这个问题，我们提出了一种高效且增强隐私的联邦学习方案PriFLRC。与现有的解决方案不同，PriFLRC可以防止私人数据泄露，即使在存在串通实体的情况下，也不需要完全信任的一方。在MNIST、CSIC 2010、SMS Spam和CIFAR-10等多个数据集上进行的实验评估表明，PriFLRC在准确率和效率方面都有较好的表现。这些结果为在现实场景中实现保护隐私的机器学习应用奠定了基础。

{"title":"PriFLRC: A secure multi-party computation-based privacy-enhanced federated learning scheme resilient to collusion","authors":"Anh-Tu Tran , The-Dung Luong , Van-Nam Huynh","doi":"10.1016/j.neucom.2025.132574","DOIUrl":"10.1016/j.neucom.2025.132574","url":null,"abstract":"<div><div>By leveraging deep learning-based technologies, a range of complex problems across various domains has been effectively addressed. However, privacy concerns render traditional centralized training methods unsuitable for sensitive data-driven scenarios, such as those in healthcare and finance. Recently, federated learning has attracted significant attention for enabling participants to collaboratively train a shared model without disclosing their local data. Nonetheless, research has shown that adversaries can still exploit the shared parameters to compromise the applications. To address this issue, we propose PriFLRC, an efficient and privacy-enhanced federated learning scheme. Unlike existing solutions, PriFLRC can prevent private data leaks even in the presence of colluding entities, without the need for a fully trusted party. Experimental evaluations conducted on multiple datasets, including MNIST, CSIC 2010, SMS Spam and CIFAR-10, demonstrate that PriFLRC excels in both accuracy and efficiency. These results lay the groundwork for implementing privacy-preserving machine learning applications in real-world scenarios.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"670 ","pages":"Article 132574"},"PeriodicalIF":6.5,"publicationDate":"2026-01-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145928299","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"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 survey of adversarial attacks on machine learning 对机器学习的对抗性攻击的调查

IF 6.5 2区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neurocomputing

Pub Date : 2026-01-02 DOI: 10.1016/j.neucom.2025.132573

Fahri Anıl Yerlikaya, Şerif Bahtiyar

Recently, machine learning has a potential to upgrade existing technology areas to a new level. In many areas like vision, security, natural language processing, and speech recognition, machine learning algorithms have been implemented to both enhance the service performance and create new services. However, these systems contain critical vulnerabilities and they are subject to adversarial attacks. In this paper, we propose a novel survey that highlights the effects of adversarial attacks during the training stage. We investigate the attacks from both the attacker and target perspectives. We divide attacks into two main categories according to their design strategies, namely poison and backdoor attacks. In each attack category, we use the targeted environment to analyze attacks with a systematic approach. Finally, we conclude the paper with challenges and future work that may help researchers to understand the anatomy of adversarial attacks. The ultimate goal of this survey is to show research directions for creating effective countermeasures against adversarial attacks on machine learning. Thus, societies will benefit from secure services that use machine learning models.

最近，机器学习有可能将现有的技术领域提升到一个新的水平。在视觉、安全、自然语言处理和语音识别等许多领域，已经实现了机器学习算法，以提高服务性能并创建新服务。然而，这些系统包含严重的漏洞，并且容易受到对抗性攻击。在本文中，我们提出了一项新的调查，突出了对抗性攻击在训练阶段的影响。我们从攻击者和目标两方面调查攻击。我们根据攻击的设计策略将攻击分为两大类，即毒杀攻击和后门攻击。在每个攻击类别中，我们使用目标环境以系统的方法分析攻击。最后，我们总结了挑战和未来的工作，这可能有助于研究人员了解对抗性攻击的解剖结构。本调查的最终目标是显示针对机器学习的对抗性攻击创建有效对策的研究方向。因此，社会将受益于使用机器学习模型的安全服务。

{"title":"A survey of adversarial attacks on machine learning","authors":"Fahri Anıl Yerlikaya, Şerif Bahtiyar","doi":"10.1016/j.neucom.2025.132573","DOIUrl":"10.1016/j.neucom.2025.132573","url":null,"abstract":"<div><div>Recently, machine learning has a potential to upgrade existing technology areas to a new level. In many areas like vision, security, natural language processing, and speech recognition, machine learning algorithms have been implemented to both enhance the service performance and create new services. However, these systems contain critical vulnerabilities and they are subject to adversarial attacks. In this paper, we propose a novel survey that highlights the effects of adversarial attacks during the training stage. We investigate the attacks from both the attacker and target perspectives. We divide attacks into two main categories according to their design strategies, namely poison and backdoor attacks. In each attack category, we use the targeted environment to analyze attacks with a systematic approach. Finally, we conclude the paper with challenges and future work that may help researchers to understand the anatomy of adversarial attacks. The ultimate goal of this survey is to show research directions for creating effective countermeasures against adversarial attacks on machine learning. Thus, societies will benefit from secure services that use machine learning models.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"670 ","pages":"Article 132573"},"PeriodicalIF":6.5,"publicationDate":"2026-01-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145928300","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"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

Safe multi-view graph convolutional network for semi-supervised classification 用于半监督分类的安全多视图卷积网络

IF 6.5 2区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neurocomputing

Pub Date : 2026-01-02 DOI: 10.1016/j.neucom.2025.132570

Dan Xiang , Xin Zhong , Hao Sun , Boxuan Tan , Pan Gao , Jinwen Zhang , Jing Ling , Haihua Du , Naiyao Liang

Graph Convolutional Network (GCN) is widely used in multi-view semi-supervised learning for its ability to capture structural and relational information. However, as the number of views increases, existing GCN-based methods often suffer from noise and inconsistencies while fusing multi-view information, leading to performance degradation. To address this issue, we propose a novel GCN-based method called Safe Multi-view Semi-supervised GCN (SMSGCN), which mitigates the risk of performance degradation caused by an increase. Our method integrates a reconstruction objective with Laplacian embedding and a safe mechanism into a unified GCN-based framework. Specifically, we use a reconstruction error based on Laplacian embedding to capture cross-view complementarity, and a safe module that dynamically adjusts safe coefficients to emphasize informative views and suppress noisy newly increased views, preventing performance degradation as the number of views increases. As a result, it can adaptively select informative views while suppressing noisy ones, thereby ensuring stable performance. In addition, we define safety from the perspective of empirical classification risk and theoretically prove that our method can achieve empirically safe multi-view semi-supervised classification. Extensive experiments conducted on multiple public benchmark datasets validate the effectiveness, robustness, and superiority of the proposed method in achieving safe multi-view semi-supervised classification.

图卷积网络（GCN）以其捕获结构信息和关系信息的能力被广泛应用于多视图半监督学习中。然而，随着视图数量的增加，现有的基于gcn的方法在融合多视图信息时经常受到噪声和不一致性的影响，导致性能下降。为了解决这个问题，我们提出了一种新的基于GCN的方法，称为安全多视图半监督GCN (SMSGCN)，它降低了由于增加而导致的性能下降的风险。我们的方法将重建目标与拉普拉斯嵌入和安全机制集成到一个统一的基于gcn的框架中。具体来说，我们使用基于拉普拉斯嵌入的重建误差来捕获交叉视图互补性，并使用安全模块动态调整安全系数来强调信息视图并抑制噪声新增加的视图，从而防止随着视图数量的增加而性能下降。因此，它可以自适应地选择信息视图，同时抑制噪声视图，从而保证稳定的性能。此外，我们从经验分类风险的角度对安全性进行了定义，并从理论上证明了我们的方法可以实现经验安全的多视图半监督分类。在多个公共基准数据集上进行的大量实验验证了该方法在实现安全的多视图半监督分类方面的有效性、鲁棒性和优越性。

{"title":"Safe multi-view graph convolutional network for semi-supervised classification","authors":"Dan Xiang , Xin Zhong , Hao Sun , Boxuan Tan , Pan Gao , Jinwen Zhang , Jing Ling , Haihua Du , Naiyao Liang","doi":"10.1016/j.neucom.2025.132570","DOIUrl":"10.1016/j.neucom.2025.132570","url":null,"abstract":"<div><div>Graph Convolutional Network (GCN) is widely used in multi-view semi-supervised learning for its ability to capture structural and relational information. However, as the number of views increases, existing GCN-based methods often suffer from noise and inconsistencies while fusing multi-view information, leading to performance degradation. To address this issue, we propose a novel GCN-based method called Safe Multi-view Semi-supervised GCN (SMSGCN), which mitigates the risk of performance degradation caused by an increase. Our method integrates a reconstruction objective with Laplacian embedding and a safe mechanism into a unified GCN-based framework. Specifically, we use a reconstruction error based on Laplacian embedding to capture cross-view complementarity, and a safe module that dynamically adjusts safe coefficients to emphasize informative views and suppress noisy newly increased views, preventing performance degradation as the number of views increases. As a result, it can adaptively select informative views while suppressing noisy ones, thereby ensuring stable performance. In addition, we define safety from the perspective of empirical classification risk and theoretically prove that our method can achieve empirically safe multi-view semi-supervised classification. Extensive experiments conducted on multiple public benchmark datasets validate the effectiveness, robustness, and superiority of the proposed method in achieving safe multi-view semi-supervised classification.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"670 ","pages":"Article 132570"},"PeriodicalIF":6.5,"publicationDate":"2026-01-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145928160","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"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

Mixture-of-experts-based hierarchical dynamic multimodal fusion network for dermatological diagnosis 基于专家混合的皮肤病诊断层次动态多模态融合网络

IF 6.5 2区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neurocomputing

Pub Date : 2026-01-02 DOI: 10.1016/j.neucom.2025.132585

Chen Hu , Feng Li , Min Li , Enguang Zuo , Xiaoyi Lv , Shumei Bao , Chengwei Rao , Xiaodong Xie , Cheng Chen , Chen Chen

In clinical computer-aided diagnosis of skin cancer, integrating skin lesion images with clinical metadata is crucial for improving diagnostic accuracy. However, most existing fusion methods rely on fixed strategies, such as weighted averaging, which do not adaptively assess the strength of feature expression. This results in an imbalance in the fusion of key discriminative information across multiple modalities and scales. To address these limitations, we propose the Mixture-of-Experts-based Hierarchical Dynamic Multimodal Fusion Network (MHDF-Net), which enables progressive and dynamic fusion of images and metadata. Specifically, we design a Local Cross-Modal Mixture-of-Experts (LC-MoE) that incorporates a Top-k Cross-Attention (TKCA) parallel branch structure, in combination with the Local Cross-Modal Gate (LC-Gate). This design enhances discriminative feature extraction across modalities and resolves the issue of dynamic complementary balance in highly similar samples. Additionally, we propose the Global Multi-scale Mixture-of-Experts (GM-MoE), which employs a multi-scale hierarchical expert architecture to adaptively integrate global contextual information of the lesion. This approach accommodates the diversity of lesion morphology by modeling spatial semantics. Extensive experiments on two publicly available skin cancer diagnosis datasets show that our model significantly outperforms existing dermatological disease classification algorithms, offering new insights for multimodal fusion in skin cancer diagnosis. The codes are publicly available at https://github.com/ChenHu-0413/MHDF-Net.

在临床皮肤癌计算机辅助诊断中，将皮肤病变图像与临床元数据相结合是提高诊断准确率的关键。然而，大多数现有的融合方法依赖于固定的策略，如加权平均，不能自适应地评估特征表达的强度。这导致了关键判别信息在多模态和尺度融合中的不平衡。为了解决这些限制，我们提出了基于混合专家的分层动态多模态融合网络（MHDF-Net），它可以实现图像和元数据的渐进和动态融合。具体来说，我们设计了一个局部跨模态混合专家（LC-MoE），它结合了Top-k交叉注意（TKCA）并行分支结构，并结合了局部跨模态门（LC-Gate）。该设计增强了跨模态的判别特征提取，并解决了高度相似样本中的动态互补平衡问题。此外，我们还提出了Global Multi-scale Mixture-of-Experts （GM-MoE）算法，该算法采用多尺度分层专家架构自适应整合病变的全局上下文信息。该方法通过空间语义建模来适应病变形态的多样性。在两个公开可用的皮肤癌诊断数据集上进行的大量实验表明，我们的模型显著优于现有的皮肤病分类算法，为皮肤癌诊断中的多模态融合提供了新的见解。这些代码可在https://github.com/ChenHu-0413/MHDF-Net上公开获取。

{"title":"Mixture-of-experts-based hierarchical dynamic multimodal fusion network for dermatological diagnosis","authors":"Chen Hu , Feng Li , Min Li , Enguang Zuo , Xiaoyi Lv , Shumei Bao , Chengwei Rao , Xiaodong Xie , Cheng Chen , Chen Chen","doi":"10.1016/j.neucom.2025.132585","DOIUrl":"10.1016/j.neucom.2025.132585","url":null,"abstract":"<div><div>In clinical computer-aided diagnosis of skin cancer, integrating skin lesion images with clinical metadata is crucial for improving diagnostic accuracy. However, most existing fusion methods rely on fixed strategies, such as weighted averaging, which do not adaptively assess the strength of feature expression. This results in an imbalance in the fusion of key discriminative information across multiple modalities and scales. To address these limitations, we propose the Mixture-of-Experts-based Hierarchical Dynamic Multimodal Fusion Network (MHDF-Net), which enables progressive and dynamic fusion of images and metadata. Specifically, we design a Local Cross-Modal Mixture-of-Experts (LC-MoE) that incorporates a Top-k Cross-Attention (TKCA) parallel branch structure, in combination with the Local Cross-Modal Gate (LC-Gate). This design enhances discriminative feature extraction across modalities and resolves the issue of dynamic complementary balance in highly similar samples. Additionally, we propose the Global Multi-scale Mixture-of-Experts (GM-MoE), which employs a multi-scale hierarchical expert architecture to adaptively integrate global contextual information of the lesion. This approach accommodates the diversity of lesion morphology by modeling spatial semantics. Extensive experiments on two publicly available skin cancer diagnosis datasets show that our model significantly outperforms existing dermatological disease classification algorithms, offering new insights for multimodal fusion in skin cancer diagnosis. The codes are publicly available at <span><span>https://github.com/ChenHu-0413/MHDF-Net</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"671 ","pages":"Article 132585"},"PeriodicalIF":6.5,"publicationDate":"2026-01-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145980736","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"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