Neurocomputing最新文献_第4页

Spatio-temporal tree attention network for forecasting traffic flow 基于时空树注意力网络的交通流预测

IF 6.5 2区计算机科学

Neurocomputing Pub Date : 2026-05-01 Epub Date: 2026-02-11 DOI: 10.1016/j.neucom.2026.133032

Haoran Li , Zhiqiang Lv , Zhaobin Ma , Jianbo Li , Xiaolong Ma , Dongxin Sun , Kangxin Guo , Jun Liu

{"title":"Spatio-temporal tree attention network for forecasting traffic flow","authors":"Haoran Li , Zhiqiang Lv , Zhaobin Ma , Jianbo Li , Xiaolong Ma , Dongxin Sun , Kangxin Guo , Jun Liu","doi":"10.1016/j.neucom.2026.133032","DOIUrl":"10.1016/j.neucom.2026.133032","url":null,"abstract":"<div><div>The Intelligent Transport Systems represent a pivotal research area within the broader context of smart city construction. It constitutes a vital component of the contemporary urban transport system, with the potential to facilitate high-quality development. The prediction of traffic flow represents a significant research area within the field of ITS. It offers a valuable opportunity to develop a robust data foundation for the planning and optimisation of urban traffic road networks. The majority of studies in this field currently employ static graphs and graph neural networks to complete the traffic flow prediction task. The use of static graphs for traffic flow prediction is inadequate for capturing the dynamic spatial and temporal characteristics of the traffic network structure. Furthermore, graph neural networks are only capable of performing local spatial characteristic analysis. To address the issue of global feature analysis of traffic network topology, multi-layer graph neural networks are required for iterative computation. The number of layers of graph neural networks increases in line with the number of nodes in the traffic network. To address the aforementioned issues, this study proposes a neural network architecture that employs a tree structure for attention computation, namely the Spatio-temporal Tree Attention Network (STTAT). In particular, this study proposes a tree-structured representation of traffic network topology and a tree-structured attention computation method for learning global features of traffic network topology. The proposed model has been evaluated on several real-world traffic datasets, and its performance has been compared with that of several baseline models. The results demonstrate that the proposed model significantly outperforms the baseline models in terms of prediction accuracy.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133032"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386719","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

ARFNet: Scale-aware adaptive receptive field network for time series forecasting ARFNet：用于时间序列预测的尺度感知自适应接受场网络

IF 6.5 2区计算机科学

Neurocomputing Pub Date : 2026-05-01 Epub Date: 2026-02-11 DOI: 10.1016/j.neucom.2026.133044

Chongyun Qin, Zhenpeng Wu, Yiting Shi, Jianliang Gao

{"title":"ARFNet: Scale-aware adaptive receptive field network for time series forecasting","authors":"Chongyun Qin, Zhenpeng Wu, Yiting Shi, Jianliang Gao","doi":"10.1016/j.neucom.2026.133044","DOIUrl":"10.1016/j.neucom.2026.133044","url":null,"abstract":"<div><div>Time series forecasting plays a crucial role in decision-making across various scenarios, including weather, traffic, and finance. Real-world time series typically exhibit complex patterns where multiple periods and dynamic trends overlap. However, traditional models based on fixed receptive fields struggle to adapt flexibly to scale variations in periodic and trend features across different application scenarios, thereby limiting their generalization capabilities. In this paper, we propose a scale-aware <strong>A</strong>daptive <strong>R</strong>eceptive <strong>F</strong>ield <strong>Net</strong>work (<strong>ARFNet</strong>). Specifically, we propose an Adaptive Receptive Field Pyramid module that dynamically adjusts the receptive field size based on the inherent characteristics of the data, providing high-quality periodic and trend feature inputs for subsequent analysis. In addition, we design a Scale-Aware Feature Synthesizer module to capture the dependencies between features at different scales, enhancing the model’s comprehension and utilization of multi-scale temporal information. Extensive experimental results on real-world datasets show that ARFNet outperforms state-of-the-art methods in time series forecasting.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133044"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386722","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

Learning from multi-view fragments: An adaptive consistency distillation framework for occluded person re-identification 从多视图片段中学习：一个自适应一致性蒸馏框架用于闭塞的人再识别

IF 6.5 2区计算机科学

Neurocomputing Pub Date : 2026-05-01 Epub Date: 2026-02-11 DOI: 10.1016/j.neucom.2026.133015

Jianfeng Dong , Shengwei Tian , Long Yu , Hongfeng You , Qimeng Yang , Jinmiao Song , Xinjun Pei , Feng Shi , Kun Wu

{"title":"Learning from multi-view fragments: An adaptive consistency distillation framework for occluded person re-identification","authors":"Jianfeng Dong , Shengwei Tian , Long Yu , Hongfeng You , Qimeng Yang , Jinmiao Song , Xinjun Pei , Feng Shi , Kun Wu","doi":"10.1016/j.neucom.2026.133015","DOIUrl":"10.1016/j.neucom.2026.133015","url":null,"abstract":"<div><div>Occluded person re-identification (ReID) faces an “information incompleteness paradox”: a single occluded view misses discriminative cues and yields ambiguous representations, while exploiting multi-view observations typically requires multi-branch inference with high computational cost. To address this dilemma, we propose <strong>Multi-View Consistency Distillation (MVCD)</strong>, a framework formulated under the <strong>Learning Using Privileged Information (LUPI)</strong> paradigm. Specifically, we construct a <em>training-only</em> teacher that has privileged access to multi-view fragments and identity annotations (available only during training), and transfer this privileged knowledge to a standard single-view student through consistency distillation. The teacher contains three training-only mechanisms: (1) <strong>Saliency-Guided Feature Purification (SGFP)</strong> to suppress occlusion-induced noise with label guidance; (2) <strong>Cross-View Patch Alignment (CVPA)</strong> to exploit patch correspondences for spatial rectification across views; and (3) <strong>Reliability-Guided Aggregation (RGA)</strong> to produce a low-variance, reliable supervision target. Crucially, all auxiliary components are discarded after training, enabling the student to recover more complete representations from occluded inputs with <strong>zero extra inference-time cost</strong>. Extensive experiments on five benchmarks show consistent improvements over strong baselines. On Occluded-DukeMTMC, MVCD achieves 70.4% Rank-1 accuracy and runs at 25 ms per image, outperforming prior state-of-the-art methods.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133015"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386736","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

Self2Rolling: Self-supervised denoising using detail-guided mask with updates Self2Rolling：使用带有更新的细节引导掩模进行自监督去噪

IF 6.5 2区计算机科学

Neurocomputing Pub Date : 2026-05-01 Epub Date: 2026-02-12 DOI: 10.1016/j.neucom.2026.132993

Bolin Song , Yiyang Wang , Changze Zhou

{"title":"Self2Rolling: Self-supervised denoising using detail-guided mask with updates","authors":"Bolin Song , Yiyang Wang , Changze Zhou","doi":"10.1016/j.neucom.2026.132993","DOIUrl":"10.1016/j.neucom.2026.132993","url":null,"abstract":"<div><div>Self-supervised learning denoisers utilizing blind spot networks have garnered significant attention due to their capability to learn a denoiser that solely relies on single noisy images. However, existing approaches employ a uniform masking strategy, wherein all pixels are masked indiscriminately, resulting in the loss of intricate details. In this paper, we introduce a novel technique for generating a detail-guided mask that is not uniformly sampled across different regions, to mitigate the loss of intricate details and enhance overall denoising quality. Furthermore, we propose a novel framework named Self2Rolling to enhance the precision of detail position guidance and prioritize reinforcing crucial yet overlooked details that require preservation. The proposed framework can be viewed as an integration of a noise measurement model with self-supervised learning techniques, exhibiting continuous enhancements in denoising performance across iterations. Extensive experiments validate the superiority of our approach compared to state-of-the-art methods. The executable code will be released upon acceptance of the paper.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 132993"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386365","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 triple-inertial accelerated alternating optimization method for deep learning training 深度学习训练的三惯性加速交替优化方法

IF 6.5 2区计算机科学

Neurocomputing Pub Date : 2026-05-01 Epub Date: 2026-02-10 DOI: 10.1016/j.neucom.2026.132997

Chengcheng Yan, Jiawei Xu, Qingsong Wang, Zheng Peng

{"title":"A triple-inertial accelerated alternating optimization method for deep learning training","authors":"Chengcheng Yan, Jiawei Xu, Qingsong Wang, Zheng Peng","doi":"10.1016/j.neucom.2026.132997","DOIUrl":"10.1016/j.neucom.2026.132997","url":null,"abstract":"<div><div>The stochastic gradient descent (SGD) algorithm has achieved remarkable success in training deep learning models. However, it has several limitations, including susceptibility to vanishing gradients, sensitivity to input data, and a lack of robust theoretical guarantees. In recent years, alternating minimization (AM) methods have emerged as a promising alternative for model training by employing gradient-free approaches to iteratively update model parameters. Despite their potential, these methods often exhibit slow convergence rates. To address this challenge, we propose a novel Triple-Inertial Accelerated Alternating Minimization (TIAM) framework for neural network training. The TIAM approach incorporates a triple-inertial acceleration strategy with a specialized approximation method, facilitating targeted acceleration of different terms in each sub-problem optimization. This integration improves the efficiency of convergence, achieving better performance with fewer iterations. Additionally, we provide a convergence analysis of the TIAM algorithm, including its global convergence properties and convergence rate. Extensive experiments validate the effectiveness of the TIAM method, demonstrating improvements in generalization capability and computational efficiency compared to existing approaches, particularly when applied to the rectified linear unit (ReLU) and its variants.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 132997"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386363","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

Learning high-order user-item relation via hyperedge for recommender system 基于hyperedge的推荐系统高阶用户-物品关系学习

IF 6.5 2区计算机科学

Neurocomputing Pub Date : 2026-05-01 Epub Date: 2026-02-07 DOI: 10.1016/j.neucom.2026.133004

Jingwen Wang , Yuguang Yan , Ruichu Cai , Michael K. Ng , Zhifeng Hao

{"title":"Learning high-order user-item relation via hyperedge for recommender system","authors":"Jingwen Wang , Yuguang Yan , Ruichu Cai , Michael K. Ng , Zhifeng Hao","doi":"10.1016/j.neucom.2026.133004","DOIUrl":"10.1016/j.neucom.2026.133004","url":null,"abstract":"<div><div>Recommender systems aim to find candidate items that are likely to be interesting to users based on their potential preferences. Existing methods mainly leverage user-item interaction data to learn a pairwise relation between a user and an item, or incorporate the social relations of users from a social network to model high-order relations among multiple users. However, complex relations exist not only among users but also among items, and high-order relations among users and items are vital for recommendation. For example, people buy products because of different latent reasons, which can be captured by relations involving multiple items. Such a high-order user-item relations have barely been studied in existing research. In this paper, we seek to extract high-order relations involving both users and items from interaction data and construct hyperedges to represent these relations. Specifically, we identify latent factors between users and items as the hyperedges, which is performed by matrix factorization on the interaction matrix. After that, we develop a hypergraph convolutional network based on hypergraph expansion to learn embeddings for users, items, as well as high-order relations in a joint representation space. By doing this, high-order relations involving multiple users and items are exploited to learn comprehensive representations of users and items for recommendation. The experimental results on several real-world datasets demonstrate the effectiveness of our proposed method.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133004"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386382","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

FMFNet: A Faster Multimodal Fusion Network for action recognition via efficient modality compensation 基于有效模态补偿的快速多模态融合网络

IF 6.5 2区计算机科学

Neurocomputing Pub Date : 2026-05-01 Epub Date: 2026-02-16 DOI: 10.1016/j.neucom.2026.133091

Shaocan Liu , Xingtao Wang , Penghong Wang , Ruiqin Xiong , Xiaopeng Fan

{"title":"FMFNet: A Faster Multimodal Fusion Network for action recognition via efficient modality compensation","authors":"Shaocan Liu , Xingtao Wang , Penghong Wang , Ruiqin Xiong , Xiaopeng Fan","doi":"10.1016/j.neucom.2026.133091","DOIUrl":"10.1016/j.neucom.2026.133091","url":null,"abstract":"<div><div>Skeleton-based methods have achieved remarkable success in human action recognition. However, their performance remains limited due to insufficient diversity of visual features inherently present in skeleton data. While existing multimodal approaches have demonstrated improved recognition accuracy by integrating RGB and skeleton modalities, these methods often rely on multi-stream architectures that involve numerous parameters, resulting in low inference speed. Therefore, efficient fusion strategies for RGB and skeleton modalities in multimodal frameworks deserve further exploration. In this paper, we propose a Faster Multimodal Fusion Network (FMFNet) for action recognition, which effectively integrates RGB and skeleton modalities using a modality compensation strategy. Compared to existing multi-stream multimodal methods, our approach significantly reduces model parameters while maintaining competitive recognition performance. Specifically, we first construct two types of mid-level features based on skeleton and RGB data. These mid-level features are then fused within the proposed modality compensation module to produce comprehensive representations, which are subsequently processed by a lightweight main stream to generate final predictions. Extensive experiments demonstrate that our FMFNet achieves a superior trade-off between recognition accuracy and computational efficiency on two large-scale data sets (NTU RGB+D 60 and 120). For instance, FMFNet obtains an accuracy of 95.5% on the cross-subject benchmark of NTU RGB+D 60, while being approximately 7.4<span><math><mo>×</mo></math></span> smaller and 2.3<span><math><mo>×</mo></math></span> faster than HM-CTNet, one of the current state-of-the-art methods.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133091"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386324","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

CC-mamba: Mamba-based color constancy with illumination prior-guided dynamic feature modulation and wavelet-domain attention mechanism cc -曼巴：基于曼巴的色彩恒常与光照先验引导的动态特征调制和小波域注意机制

IF 6.5 2区计算机科学

Neurocomputing Pub Date : 2026-05-01 Epub Date: 2026-02-16 DOI: 10.1016/j.neucom.2026.133068

Junyi Liu , Li Zhuo , Hui Zhang , Haokui Xu , Xiaoguang Li

{"title":"CC-mamba: Mamba-based color constancy with illumination prior-guided dynamic feature modulation and wavelet-domain attention mechanism","authors":"Junyi Liu , Li Zhuo , Hui Zhang , Haokui Xu , Xiaoguang Li","doi":"10.1016/j.neucom.2026.133068","DOIUrl":"10.1016/j.neucom.2026.133068","url":null,"abstract":"<div><div>Computational Color Constancy (CC) aims to correct color deviations in images caused by the different illumination conditions. The existing methods, based on Convolutional Neural Networks (CNNs) and Transformers, exhibit limitations in extracting global contextual information or achieving computational efficiency, leading to inaccurate illumination estimation. Therefore, in this paper, a Mamba-based CC method, named CC-Mamba, is proposed. It employs VMamba as the backbone to construct an end-to-end illumination estimation model. Specifically, a Wavelet-domain Attention (WA) mechanism is first proposed, which captures local details (<em>e.g.,</em> edges and textures) and global illumination information while suppressing noise, thereby, facilitating improved illumination estimation accuracy. Next, an Illumination Prior-guided Dynamic Feature Modulation (IP-DFM) module is proposed, which utilizes illumination prior information obtained from high-level features to dynamically modulate the mid-level features, thus enhancing the model’s robustness against complex illumination variations and challenging scenarios. Finally, the Multi-Scale Gated Attentional Feature Fusion (MS-GAFF) module is designed, which employs a gating mechanism for illumination-related feature selection, coupled with a channel-spatial dual-dimensional dynamic weighting strategy to adaptively weigh the features at different scales to promote the representation ability of the features. Experimental results on the NUS 8-Camera and CCD (Color Checker Dataset) benchmark datasets demonstrate that CC-Mamba achieves superior illumination estimation accuracy compared to the state-of-the-art methods. Furthermore, the corrected image has a better perceptual quality.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133068"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386373","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

An optimal control method for uncertain process industry based on working condition relevance and policy transfer 基于工况关联和策略转移的不确定过程工业最优控制方法

IF 6.5 2区计算机科学

Neurocomputing Pub Date : 2026-05-01 Epub Date: 2026-02-11 DOI: 10.1016/j.neucom.2026.133028

Can Zhou, Xuan Ouyang, Hongqiu Zhu, Tianhao Liu

{"title":"An optimal control method for uncertain process industry based on working condition relevance and policy transfer","authors":"Can Zhou, Xuan Ouyang, Hongqiu Zhu, Tianhao Liu","doi":"10.1016/j.neucom.2026.133028","DOIUrl":"10.1016/j.neucom.2026.133028","url":null,"abstract":"<div><div>Energy consumption optimal control is a key challenge in process industries and is essential for enhancing production efficiency and reducing costs. However, the intricate physical and chemical reaction mechanisms in process industries make it difficult to establish precise process models. Moreover, unpredictable fluctuations in raw materials introduce uncertain variations in working conditions, causing unknown working conditions to emerge. Under this circumstance, the mismatch between the trained controller and the unknown working condition leads to deviation in control results. To get around these issues, a control method based on transfer reinforcement learning is proposed, which comprises two constituents. First, multiple RL controllers are trained under different working conditions to learn their distinct characteristics independently. By embedding domain knowledge into the reward function, the learning of control policies is guided to optimize energy consumption. Second, seeing that retraining RL controllers when encountering unknown working conditions is a costly endeavor, the concept of working condition relevance is proposed to quantify the degree of correlation between known and unknown working conditions. By utilizing a small amount of interaction experience with unknown and known working conditions, the control policies for known working conditions are transferred to the ones for unknown working conditions. As a result, the learning costs of the control policy are reduced while the negative transfer is mitigated to a certain extent. The simulation experiments on the zinc electrowinning process demonstrate that our method transfers control policies effectively when unknown working conditions arise, reducing learning costs noticeably.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133028"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386733","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

MetaGT: A lightweight graph transformer via meta learning for large-scale graphs MetaGT：一个轻量级的图转换器，通过元学习来处理大规模的图

IF 6.5 2区计算机科学

Neurocomputing Pub Date : 2026-05-01 Epub Date: 2026-02-12 DOI: 10.1016/j.neucom.2026.132992

Wenting Wang , Yunfei Tian , Hailong Zhang

{"title":"MetaGT: A lightweight graph transformer via meta learning for large-scale graphs","authors":"Wenting Wang , Yunfei Tian , Hailong Zhang","doi":"10.1016/j.neucom.2026.132992","DOIUrl":"10.1016/j.neucom.2026.132992","url":null,"abstract":"<div><div>Graph Transformers (GTs) excel at modeling long-distance dependencies and global information in graphs, but their scalability on large-scale graphs is limited by the quadratic growth of memory and computation costs introduced by self-attention. To address these challenges, existing methods such as linear attention mechanisms and graph partitioning have been proposed. However, these methods have inherent limitations. Linear attention mechanisms still incur significant computational overhead, while subgraph-based training often leads to reduced accuracy due to the lack of global context and the structural differences among subgraphs. In this paper, we propose MetaGT, a lightweight Graph Transformer model for large-scale graphs. MetaGT first partitions the global graph into structurally compact subgraphs to reduce memory requirements. To further ensure both efficiency and accuracy, we introduce a “pre- and post-training” paradigm into subgraph-based GT models. In pre-training, meta-learning is employed on a subset of subgraphs to obtain generalizable parameters, which are then enhanced by minimal fine-tuning during post-training. MetaGT is plug-and-play and can be seamlessly integrated into mainstream GT models. Extensive experiments demonstrate that MetaGT consistently achieves promising efficiency, memory consumption, and accuracy across various backbone GT models on large-scale graph datasets. Our code is available at <span><span>https://anonymous.4open.science/r/MetaGT-D2B0</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 132992"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386735","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