Neural Networks最新文献_第9页

SAGCN: A syntactic aware multi-branch graph attention network with structural bias for aspect sentiment triplet extraction 面向面向情感三元组提取的结构偏差多分支图注意网络。

IF 6.3 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neural Networks

Pub Date : 2026-06-01 Epub Date: 2026-01-13 DOI: 10.1016/j.neunet.2026.108596

Xin Xiao , Bin Gao , Zelong Su , Linlin Li , Shutian Liu , Zhengjun Liu

Aspect-based sentiment triplet extraction (ASTE), a newly developed and complex subtask within aspect-based sentiment analysis, focuses on detecting aspect terms, opinion terms, and establishing sentiment polarity from human language, thereby extracting triplets composed of these three elements. Although numerous methods have been developed in previous research to tackle this task, these ASTE methods exhibit weak interactions in constructing contextual representations and overlook the syntactic relationships between aspect terms and opinion terms. Therefore, this paper proposes a syntax-aware multi-branch graph attention network to address this issue. We have designed an efficient approach that integrates new structural biases into pre-trained language models through adapters to enhance the original mappings in self-attention, significantly reducing the parameter requirements and achieving lower latency. Simultaneously, we have devised a syntax-aware attention mechanism that not only discerns edges with varying dependency types as well as those with identical types, learning the representation of each edge in the graph relying on the dependency types of neighboring edges, thereby enabling more accurate graph propagation. Finally, we have designed a special fusion interaction layer that achieves the final text representation by merging different branch features with varying weights. Through a range of tests performed on four widely accessible datasets, it was demonstrated that the introduction of structural bias adapters is both effective and efficient. The proposed method improved the average F1 score by up to 4.11% compared to all baseline models, while also exhibiting good interpretability. Additionally, the experimental results highlighted the robustness and effectiveness of SAGCN, significantly outperforming the compared state-of-the-art baseline models.

基于方面的情感三联体提取（ASTE）是基于方面的情感分析中一个新发展的复杂子任务，其重点是从人类语言中检测方面术语、观点术语和建立情感极性，从而提取由这三个元素组成的三联体。尽管在之前的研究中已经开发了许多方法来解决这一任务，但这些方法在构建上下文表示时表现出弱交互，并且忽略了方面术语和意见术语之间的句法关系。为此，本文提出了一种语法感知的多分支图关注网络来解决这一问题。我们设计了一种有效的方法，通过适配器将新的结构偏差集成到预训练的语言模型中，以增强自注意中的原始映射，显著降低参数要求并实现更低的延迟。同时，我们设计了一种语法感知的注意机制，不仅可以区分不同依赖类型的边和相同类型的边，还可以根据相邻边的依赖类型学习图中每条边的表示，从而实现更准确的图传播。最后，我们设计了一个特殊的融合交互层，通过合并不同权重的分支特征来实现最终的文本表示。通过在四个可广泛访问的数据集上进行的一系列测试，证明了引入结构偏差适配器既有效又高效。与所有基线模型相比，该方法将F1平均得分提高了4.11%，同时也具有良好的可解释性。此外，实验结果突出了SAGCN的鲁棒性和有效性，显著优于最先进的基线模型。

{"title":"SAGCN: A syntactic aware multi-branch graph attention network with structural bias for aspect sentiment triplet extraction","authors":"Xin Xiao , Bin Gao , Zelong Su , Linlin Li , Shutian Liu , Zhengjun Liu","doi":"10.1016/j.neunet.2026.108596","DOIUrl":"10.1016/j.neunet.2026.108596","url":null,"abstract":"<div><div>Aspect-based sentiment triplet extraction (ASTE), a newly developed and complex subtask within aspect-based sentiment analysis, focuses on detecting aspect terms, opinion terms, and establishing sentiment polarity from human language, thereby extracting triplets composed of these three elements. Although numerous methods have been developed in previous research to tackle this task, these ASTE methods exhibit weak interactions in constructing contextual representations and overlook the syntactic relationships between aspect terms and opinion terms. Therefore, this paper proposes a syntax-aware multi-branch graph attention network to address this issue. We have designed an efficient approach that integrates new structural biases into pre-trained language models through adapters to enhance the original mappings in self-attention, significantly reducing the parameter requirements and achieving lower latency. Simultaneously, we have devised a syntax-aware attention mechanism that not only discerns edges with varying dependency types as well as those with identical types, learning the representation of each edge in the graph relying on the dependency types of neighboring edges, thereby enabling more accurate graph propagation. Finally, we have designed a special fusion interaction layer that achieves the final text representation by merging different branch features with varying weights. Through a range of tests performed on four widely accessible datasets, it was demonstrated that the introduction of structural bias adapters is both effective and efficient. The proposed method improved the average F1 score by up to 4.11% compared to all baseline models, while also exhibiting good interpretability. Additionally, the experimental results highlighted the robustness and effectiveness of SAGCN, significantly outperforming the compared state-of-the-art baseline models.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"198 ","pages":"Article 108596"},"PeriodicalIF":6.3,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146031529","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

FRM-PTQ: Feature relationship matching enhanced low-bit post-training quantization for large language models FRM-PTQ：针对大型语言模型的特征关系匹配增强的低比特训练后量化。

IF 6.3 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neural Networks

Pub Date : 2026-06-01 Epub Date: 2026-01-17 DOI: 10.1016/j.neunet.2026.108619

Chao Zeng , Jiaqi Zhao , Miao Zhang , Li Wang , Weili Guan , Liqiang Nie

Post-Training Quantization (PTQ) has emerged as an effective approach to reduce memory and computational demands during LLMs inference. However, existing PTQ methods are highly sensitive to ultra-low-bit quantization with significant performance loss, which is further exacerbated by recently released advanced models like LLaMA-3 and LLaMA-3.1. To address this challenge, we propose a novel PTQ framework, termed FRM-PTQ, by introducing feature relationship matching. This approach integrates token-level relationship modeling and structure-level distribution alignment based on the intra-block self-distillation framework to effectively mitigate significant performance degradation caused by low-bit quantization. Unlike conventional MSE loss methods, which focus solely on point-to-point discrepancies, feature relationship matching captures feature representations in high-dimensional spaces to effectively bridge the representation gap between quantized and full-precision blocks. Additionally, we propose a multi-granularity per-group quantization technique featuring a customized kernel, designed based on the quantization sensitivity of decoder block, to further relieve the quantization performance degradation. Extensive experimental results demonstrate that our method achieves outstanding performance in the W4A4 low-bit scenario, maintaining near full-precision accuracy while delivering a 2 ×  throughput improvement and a 3.17 ×  memory reduction. This advantage is particularly evident in the latest models such as LLaMA-3, LLaMA-3.1 and Qwen2.5 models, as well as in the W3A3 extreme low-bit scenarios. Codes are available at https://github.com/HITSZ-Miao-Group/FRM.

训练后量化（PTQ）已成为llm推理过程中减少内存和计算需求的有效方法。然而，现有的PTQ方法对超低比特量化非常敏感，性能损失很大，最近发布的先进模型如LLaMA-3和LLaMA-3.1进一步加剧了这一问题。为了解决这一挑战，我们提出了一个新的PTQ框架，称为FRM-PTQ，通过引入特征关系匹配。该方法结合了基于块内自蒸馏框架的令牌级关系建模和结构级分布对齐，有效缓解了低比特量化导致的性能下降。与仅关注点对点差异的传统MSE损失方法不同，特征关系匹配捕获高维空间中的特征表示，有效地弥合了量化块和全精度块之间的表示差距。此外，我们提出了一种基于译码块量化灵敏度的定制内核的多粒度每组量化技术，以进一步缓解量化性能的下降。大量的实验结果表明，我们的方法在W4A4低比特场景中取得了出色的性能，在提供2 × 吞吐量改进和3.17 × 内存减少的同时，保持了接近全精度的精度。这一优势在最新型号如LLaMA-3、LLaMA-3.1和Qwen2.5型号以及W3A3极低比特场景中尤为明显。代码可在https://github.com/HITSZ-Miao-Group/FRM上获得。

{"title":"FRM-PTQ: Feature relationship matching enhanced low-bit post-training quantization for large language models","authors":"Chao Zeng , Jiaqi Zhao , Miao Zhang , Li Wang , Weili Guan , Liqiang Nie","doi":"10.1016/j.neunet.2026.108619","DOIUrl":"10.1016/j.neunet.2026.108619","url":null,"abstract":"<div><div>Post-Training Quantization (PTQ) has emerged as an effective approach to reduce memory and computational demands during LLMs inference. However, existing PTQ methods are highly sensitive to ultra-low-bit quantization with significant performance loss, which is further exacerbated by recently released advanced models like LLaMA-3 and LLaMA-3.1. To address this challenge, we propose a novel PTQ framework, termed <strong>FRM-PTQ</strong>, by introducing feature relationship matching. This approach integrates token-level relationship modeling and structure-level distribution alignment based on the intra-block self-distillation framework to effectively mitigate significant performance degradation caused by low-bit quantization. Unlike conventional MSE loss methods, which focus solely on point-to-point discrepancies, feature relationship matching captures feature representations in high-dimensional spaces to effectively bridge the representation gap between quantized and full-precision blocks. Additionally, we propose a multi-granularity per-group quantization technique featuring a customized kernel, designed based on the quantization sensitivity of decoder block, to further relieve the quantization performance degradation. Extensive experimental results demonstrate that our method achieves outstanding performance in the W4A4 low-bit scenario, maintaining near full-precision accuracy while delivering a 2 ×  throughput improvement and a 3.17 ×  memory reduction. This advantage is particularly evident in the latest models such as LLaMA-3, LLaMA-3.1 and Qwen2.5 models, as well as in the W3A3 extreme low-bit scenarios. Codes are available at <span><span>https://github.com/HITSZ-Miao-Group/FRM</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"198 ","pages":"Article 108619"},"PeriodicalIF":6.3,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146047368","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

AttCo: Attention-based co-Learning fusion of deep feature representation for medical image segmentation using multimodality 基于关注的深度特征表示融合多模态医学图像分割

IF 6.3 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neural Networks

Pub Date : 2026-06-01 Epub Date: 2026-01-08 DOI: 10.1016/j.neunet.2026.108572

Duy-Phuong Dao , Hyung-Jeong Yang , Soo-Hyung Kim , Sae-Ryung Kang

Accurate tissue segmentation is crucial for advancing healthcare, particularly in disease prediction and treatment planning. Precisely identifying abnormal tissue locations is a critical step for clinical analysis. While medical image segmentation increasingly utilizes multimodal and three-dimensional (3D) information to capture spatial relationships, current methods often struggle to effectively learn complementary information from multiple inputs, especially with complex 3D structures. In this study, we introduce AttCo, a novel multimodal semantic segmentation network built upon an attention-based co-learning fusion of deep feature representations. AttCo first employs multiple encoder branches to extract unimodal 3D representations from each imaging modality. These unimodal representations are then processed by a co-learning fusion module, which integrates both intra-modality (using SEAT) and inter-modality (using OSCAT) feature learning components. This dual approach ensures the capture of intricate interactions within each modality and across different modalities. Finally, these fused features pass through an up-sampling module to generate 3D segmented tumor maps. Our end-to-end deep network effectively addresses two key aspects: (i) extracting robust unimodal 3D representations and (ii) exploiting comprehensive inter- and intra-modality feature interactions. Experimental results demonstrate that AttCo significantly outperforms competing methods in terms of Dice score across various datasets. The source code can be found here https://github.com/duyphuongcri/AttCo.

准确的组织分割对于推进医疗保健至关重要，特别是在疾病预测和治疗计划方面。准确识别异常组织位置是临床分析的关键步骤。虽然医学图像分割越来越多地利用多模态和三维（3D）信息来捕捉空间关系，但目前的方法往往难以有效地从多个输入中学习互补信息，特别是复杂的3D结构。在这项研究中，我们引入了一种新的多模态语义分割网络AttCo，它建立在基于注意力的深度特征表征融合的共同学习基础上。AttCo首先使用多个编码器分支从每个成像模态中提取单模态3D表示。这些单模态表示然后由共同学习融合模块处理，该模块集成了模态内（使用SEAT）和模态间（使用OSCAT）特征学习组件。这种双重方法确保捕获每个模态内部和不同模态之间的复杂交互。最后，这些融合的特征通过上采样模块生成三维分割的肿瘤图。我们的端到端深度网络有效地解决了两个关键方面：(i)提取稳健的单模态3D表示和（ii）利用全面的模态间和模态内特征相互作用。实验结果表明，AttCo在不同数据集的Dice得分方面明显优于竞争方法。源代码可以在这里找到https://github.com/duyphuongcri/AttCo。

{"title":"AttCo: Attention-based co-Learning fusion of deep feature representation for medical image segmentation using multimodality","authors":"Duy-Phuong Dao , Hyung-Jeong Yang , Soo-Hyung Kim , Sae-Ryung Kang","doi":"10.1016/j.neunet.2026.108572","DOIUrl":"10.1016/j.neunet.2026.108572","url":null,"abstract":"<div><div>Accurate tissue segmentation is crucial for advancing healthcare, particularly in disease prediction and treatment planning. Precisely identifying abnormal tissue locations is a critical step for clinical analysis. While medical image segmentation increasingly utilizes multimodal and three-dimensional (3D) information to capture spatial relationships, current methods often struggle to effectively learn complementary information from multiple inputs, especially with complex 3D structures. In this study, we introduce AttCo, a novel multimodal semantic segmentation network built upon an attention-based co-learning fusion of deep feature representations. AttCo first employs multiple encoder branches to extract unimodal 3D representations from each imaging modality. These unimodal representations are then processed by a co-learning fusion module, which integrates both intra-modality (using SEAT) and inter-modality (using OSCAT) feature learning components. This dual approach ensures the capture of intricate interactions within each modality and across different modalities. Finally, these fused features pass through an up-sampling module to generate 3D segmented tumor maps. Our end-to-end deep network effectively addresses two key aspects: (i) extracting robust unimodal 3D representations and (ii) exploiting comprehensive inter- and intra-modality feature interactions. Experimental results demonstrate that AttCo significantly outperforms competing methods in terms of Dice score across various datasets. The source code can be found here <span><span>https://github.com/duyphuongcri/AttCo</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"198 ","pages":"Article 108572"},"PeriodicalIF":6.3,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145981716","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

FluidFormer : Transformer with continuous convolution for particle-based fluid simulation FluidFormer：具有连续卷积的变压器，用于基于颗粒的流体模拟

IF 6.3 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neural Networks

Pub Date : 2026-06-01 Epub Date: 2026-01-21 DOI: 10.1016/j.neunet.2026.108631

Nianyi Wang, Shuai Zheng, Yu Chen, Hai Zhao, Zhou Fang

Learning-based fluid simulation has emerged as an efficient alternative to traditional Navier-Stokes solvers. However, existing neural methods that build upon Smoothed Particle Hydrodynamics (SPH) predominantly rely on local particle interactions, which induces instability in complex scenarios due to error accumulation. To address this, we introduce FluidFormer, a novel architecture that establishes a hierarchical local-global modeling paradigm. The core of our model is the Fluid Attention Block (FAB), a co-design that orchestrates continuous convolution for locality with self-attention for global corrective long-range hydrodynamic phenomena. Embedded in a dual-pipeline network, our approach seamlessly fuses inductive physical biases with structured global reasoning. Extensive experiments show that FluidFormer achieves state-of-the-art performance, with significantly improved stability and generalization in challenging fluid scenes, demonstrating its potential as a robust simulator for complex physical systems.

基于学习的流体模拟已经成为传统Navier-Stokes解算器的有效替代方案。然而，现有的基于光滑粒子流体动力学（SPH）的神经方法主要依赖于局部粒子相互作用，这在复杂的情况下由于误差积累而导致不稳定。为了解决这个问题，我们介绍了FluidFormer，这是一种新的架构，可以建立分层的局部全局建模范式。我们模型的核心是流体注意块（FAB），这是一种协同设计，它协调了局部的连续卷积和全局校正远程流体动力学现象的自关注。我们的方法嵌入在双管道网络中，将归纳物理偏差与结构化全局推理无缝融合。大量实验表明，FluidFormer实现了最先进的性能，在具有挑战性的流体场景中具有显着提高的稳定性和通用性，证明了其作为复杂物理系统鲁棒模拟器的潜力。

引用次数: 0

NAR Broad Learning System for dynamical systems prediction 用于动态系统预测的NAR广义学习系统

IF 6.3 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neural Networks

Pub Date : 2026-06-01 Epub Date: 2026-01-15 DOI: 10.1016/j.neunet.2026.108617

Shuran Wang , Hua Chen , Heng Xiong , Kai Wang , Xiaogang Zhang

Dynamical systems evolve over time, and predicting their behavior is difficult because of their complex spatiotemporal relationship. Although data-driven models have achieved great success in dynamical system analysis, extracting temporal dynamic and spatial features simultaneously and maintaining fast training and updating speeds are always the bottleneck of data-driven methods. In this paper, we propose the Nonlinear Autoregression Broad Learning System (NAR-BLS), a novel shallow network for dynamical system prediction. NAR-BLS is a shallow randomized flatten network. It embeds a temporal feature capture branch into the original structure of BLS to extract temporal dynamic features of the input data, and simultaneously extracts the spatial features of the system by mapping feature nodes and enhancement nodes in a separated-aggregated way. The dynamic features and the spatial features are concatenated to the output layer for prediction. Only the weights of the output layer of NAR-BLS are computed using ridge regression. Thus, it has the excellent advantage of rapid training speed and updating ability. Experimental results on two typical chaotic systems and four real-world datasets demonstrate the superior performance of NAR-BLS.

动态系统随着时间的推移而演变，由于其复杂的时空关系，预测其行为是困难的。虽然数据驱动模型在动力系统分析中取得了巨大的成功，但同时提取时间动态特征和空间特征，保持快速的训练和更新速度一直是数据驱动方法的瓶颈。本文提出了非线性自回归广义学习系统（NAR-BLS），这是一种用于动态系统预测的新型浅层网络。NAR-BLS是一种浅层随机化平坦网络。该方法在原始BLS结构中嵌入一个时间特征捕获分支，提取输入数据的时间动态特征，同时通过对特征节点和增强节点的分离-聚合映射，提取系统的空间特征。将动态特征和空间特征连接到输出层进行预测。仅使用脊回归计算NAR-BLS输出层的权值。因此，它具有训练速度快、更新能力强的优异优势。在两个典型混沌系统和四个实际数据集上的实验结果证明了该算法的优越性能。

{"title":"NAR Broad Learning System for dynamical systems prediction","authors":"Shuran Wang , Hua Chen , Heng Xiong , Kai Wang , Xiaogang Zhang","doi":"10.1016/j.neunet.2026.108617","DOIUrl":"10.1016/j.neunet.2026.108617","url":null,"abstract":"<div><div>Dynamical systems evolve over time, and predicting their behavior is difficult because of their complex spatiotemporal relationship. Although data-driven models have achieved great success in dynamical system analysis, extracting temporal dynamic and spatial features simultaneously and maintaining fast training and updating speeds are always the bottleneck of data-driven methods. In this paper, we propose the Nonlinear Autoregression Broad Learning System (NAR-BLS), a novel shallow network for dynamical system prediction. NAR-BLS is a shallow randomized flatten network. It embeds a temporal feature capture branch into the original structure of BLS to extract temporal dynamic features of the input data, and simultaneously extracts the spatial features of the system by mapping feature nodes and enhancement nodes in a separated-aggregated way. The dynamic features and the spatial features are concatenated to the output layer for prediction. Only the weights of the output layer of NAR-BLS are computed using ridge regression. Thus, it has the excellent advantage of rapid training speed and updating ability. Experimental results on two typical chaotic systems and four real-world datasets demonstrate the superior performance of NAR-BLS.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"198 ","pages":"Article 108617"},"PeriodicalIF":6.3,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039361","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A causal bidirectional selective state space model for imaging genetics in neurodegenerative diseases 神经退行性疾病成像遗传学的因果双向选择状态空间模型

IF 6.3 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neural Networks

Pub Date : 2026-06-01 Epub Date: 2026-01-10 DOI: 10.1016/j.neunet.2026.108587

Hongrui Liu , Yuanyuan Gui , Binglei Zhao , Hui Lu , Manhua Liu

Brain imaging genetics aims to uncover the pathological mechanisms and improve the diagnosis of brain diseases, particularly neurodegenerative disorders. While deep learning has advanced feature extraction and association modeling in this field, there are still two major challenges: extracting meaningful information from long genetic sequences and establishing causal relationships among genetics, imaging, and disease. To address these challenges, this paper proposes a deep causal bidirectional selective state space model (CausalMamba) that integrates multi-level feature extraction and causal inference into a unified representation learning framework. First, the long-sequence genetic data and whole-brain imaging data are divided into localized parts, extracting the fine-grained features. Then, a causal inference strategy based on counterfactual reasoning and contrastive learning is proposed to identify the most relevant genetic and imaging features and to construct a causal chain from genetics to disease via imaging. Finally, a bidirectional selective state space model (BiMamba) efficiently integrates the selected features into modality-specific global features, enabling accurate disease diagnosis. Our model is trained jointly on genetic and imaging data, but requires only genetic data at test time. We validate the proposed method on the simulated dataset, Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset, and Parkinson’s Progression Markers Initiative (PPMI) dataset. Experimental results show that our method achieves accuracies of 80.5% and 77.3% using only genetic data in classifying Alzheimer’s and Parkinson’s diseases from normal controls, respectively, with the relative improvements of 4.2% and 2.7% over the state-of-the-art methods while also being more computationally efficient. The results demonstrate that CausalMamba can effectively identify causally relevant biomarkers across the entire genome and brain.

脑成像遗传学旨在揭示病理机制，提高脑部疾病，特别是神经退行性疾病的诊断。虽然深度学习在该领域具有先进的特征提取和关联建模，但仍然存在两个主要挑战：从长基因序列中提取有意义的信息，以及在遗传学、成像和疾病之间建立因果关系。为了解决这些挑战，本文提出了一种深度因果双向选择状态空间模型（CausalMamba），该模型将多层次特征提取和因果推理集成到统一的表征学习框架中。首先，对长序列遗传数据和全脑成像数据进行局部分割，提取细粒度特征；然后，提出了一种基于反事实推理和对比学习的因果推理策略，以识别最相关的遗传和成像特征，并通过成像构建从遗传到疾病的因果链。最后，双向选择状态空间模型（BiMamba）有效地将选择的特征集成到特定于模式的全局特征中，从而实现准确的疾病诊断。我们的模型是在遗传和成像数据上联合训练的，但只需要测试时的遗传数据。我们在模拟数据集、阿尔茨海默病神经影像学倡议（ADNI）数据集和帕金森病进展标志物倡议（PPMI）数据集上验证了所提出的方法。实验结果表明，该方法仅使用遗传数据对阿尔茨海默病和帕金森病进行分类的准确率分别为80.5%和77.3%，相对于目前的方法分别提高4.2%和2.7%，同时计算效率也更高。结果表明，CausalMamba可以有效地识别整个基因组和大脑中的因果相关生物标志物。

{"title":"A causal bidirectional selective state space model for imaging genetics in neurodegenerative diseases","authors":"Hongrui Liu , Yuanyuan Gui , Binglei Zhao , Hui Lu , Manhua Liu","doi":"10.1016/j.neunet.2026.108587","DOIUrl":"10.1016/j.neunet.2026.108587","url":null,"abstract":"<div><div>Brain imaging genetics aims to uncover the pathological mechanisms and improve the diagnosis of brain diseases, particularly neurodegenerative disorders. While deep learning has advanced feature extraction and association modeling in this field, there are still two major challenges: extracting meaningful information from long genetic sequences and establishing causal relationships among genetics, imaging, and disease. To address these challenges, this paper proposes a deep causal bidirectional selective state space model (CausalMamba) that integrates multi-level feature extraction and causal inference into a unified representation learning framework. First, the long-sequence genetic data and whole-brain imaging data are divided into localized parts, extracting the fine-grained features. Then, a causal inference strategy based on counterfactual reasoning and contrastive learning is proposed to identify the most relevant genetic and imaging features and to construct a causal chain from genetics to disease via imaging. Finally, a bidirectional selective state space model (BiMamba) efficiently integrates the selected features into modality-specific global features, enabling accurate disease diagnosis. Our model is trained jointly on genetic and imaging data, but requires only genetic data at test time. We validate the proposed method on the simulated dataset, Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset, and Parkinson’s Progression Markers Initiative (PPMI) dataset. Experimental results show that our method achieves accuracies of 80.5% and 77.3% using only genetic data in classifying Alzheimer’s and Parkinson’s diseases from normal controls, respectively, with the relative improvements of 4.2% and 2.7% over the state-of-the-art methods while also being more computationally efficient. The results demonstrate that CausalMamba can effectively identify causally relevant biomarkers across the entire genome and brain.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"198 ","pages":"Article 108587"},"PeriodicalIF":6.3,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145981628","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Learning frequency-aware graph fraud detection 学习频率感知图欺诈检测

IF 6.3 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neural Networks

Pub Date : 2026-06-01 Epub Date: 2026-01-18 DOI: 10.1016/j.neunet.2026.108600

Wei Zhao , Hao Chen

Graph Fraud Detection (GFD) has become a critical task in online systems such as financial networks, review platforms, and social media, where fraudulent behaviors are inherently rare and often embedded within benign communities. Graph Neural Networks (GNNs) have shown promise for GFD by modeling relational dependencies; however, their effectiveness is severely limited by two persistent challenges: extreme label imbalance and the coexistence of homophilic and heterophilic connections. Most existing approaches attempt to mitigate these issues by modifying graph structures or suppressing heterophilic neighbors, which may introduce bias and scalability limitations. In this work, we address GFD from a frequency-domain perspective and propose a Frequency-aware Graph Neural Network (F-GNN). The core insight is that heterophilic interactions often manifest as high-frequency signals that are diluted by dominant low-frequency homophilic patterns in conventional message passing. F-GNN explicitly decouples node representations in the graph frequency domain and employs node-adaptive spectral gating to selectively emphasize informative high-frequency components. In addition, a fraud-aware representation fusion mechanism is introduced to counteract label imbalance during neighborhood aggregation. Extensive experiments on four benchmark datasets-Yelp, Amazon, T-Finance, and T-Social-demonstrate that F-GNN consistently outperforms state-of-the-art GNN-based fraud detection methods under both supervised and semi-supervised settings, achieving up to 99.81% AUC and 96.65% F1-Macro. These results highlight the effectiveness of frequency-aware modeling as a principled alternative to structure-based heuristics for graph fraud detection.

图形欺诈检测（GFD）已经成为金融网络、评论平台和社交媒体等在线系统中的一项关键任务，在这些系统中，欺诈行为本来就很少见，而且往往嵌入良性社区。图神经网络（gnn）通过对关系依赖关系进行建模，显示出对GFD的希望；然而，他们的有效性受到两个持续挑战的严重限制：极端的标签不平衡和同性恋和异性恋关系的共存。大多数现有方法试图通过修改图结构或抑制异亲邻居来缓解这些问题，这可能会引入偏差和可扩展性限制。在这项工作中，我们从频域角度解决了GFD问题，并提出了一种频率感知图神经网络（F-GNN）。核心观点是，在传统的信息传递中，亲异性相互作用通常表现为高频信号，被占主导地位的低频亲同性模式所稀释。F-GNN显式解耦了图频域中的节点表示，并采用节点自适应频谱门控来选择性地强调信息丰富的高频成分。此外，引入欺诈感知表示融合机制来抵消邻域聚合过程中的标签不平衡。在四个基准数据集（yelp、Amazon、T-Finance和t -社交）上进行的大量实验表明，在监督和半监督设置下，F-GNN始终优于最先进的基于gnn的欺诈检测方法，达到99.81%的AUC和96.65%的F1-Macro。这些结果突出了频率感知建模作为基于结构的启发式方法在图欺诈检测方面的有效性。

{"title":"Learning frequency-aware graph fraud detection","authors":"Wei Zhao , Hao Chen","doi":"10.1016/j.neunet.2026.108600","DOIUrl":"10.1016/j.neunet.2026.108600","url":null,"abstract":"<div><div>Graph Fraud Detection (GFD) has become a critical task in online systems such as financial networks, review platforms, and social media, where fraudulent behaviors are inherently rare and often embedded within benign communities. Graph Neural Networks (GNNs) have shown promise for GFD by modeling relational dependencies; however, their effectiveness is severely limited by two persistent challenges: extreme label imbalance and the coexistence of homophilic and heterophilic connections. Most existing approaches attempt to mitigate these issues by modifying graph structures or suppressing heterophilic neighbors, which may introduce bias and scalability limitations. In this work, we address GFD from a frequency-domain perspective and propose a Frequency-aware Graph Neural Network (F-GNN). The core insight is that heterophilic interactions often manifest as high-frequency signals that are diluted by dominant low-frequency homophilic patterns in conventional message passing. F-GNN explicitly decouples node representations in the graph frequency domain and employs node-adaptive spectral gating to selectively emphasize informative high-frequency components. In addition, a fraud-aware representation fusion mechanism is introduced to counteract label imbalance during neighborhood aggregation. Extensive experiments on four benchmark datasets-Yelp, Amazon, T-Finance, and T-Social-demonstrate that F-GNN consistently outperforms state-of-the-art GNN-based fraud detection methods under both supervised and semi-supervised settings, achieving up to 99.81% AUC and 96.65% F1-Macro. These results highlight the effectiveness of frequency-aware modeling as a principled alternative to structure-based heuristics for graph fraud detection.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"198 ","pages":"Article 108600"},"PeriodicalIF":6.3,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039237","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

Deep fine-grained clustering with model reusing 具有模型重用的深细粒度集群。

IF 6.3 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neural Networks

Pub Date : 2026-06-01 Epub Date: 2026-01-15 DOI: 10.1016/j.neunet.2026.108604

Jie Hong, Xulun Ye, Jieyu Zhao

Deep clustering, which extends deep models to the clustering task, has attracted many attentions due to the high clustering performance. However, conventional deep clustering method assume that sample in different class are slightly different. Real application is complicated where observations are achieved from highly similar samples. In this paper, we study this fine-grained clustering task, while traditional coarse-grained clustering has difficulty capturing subtle semantic differences, which often leads to unclear decision boundaries between clusters of similar features. Our goal is to learn feature representations that encourage fine-grained data to form clear cluster boundaries in the embedding space. In this paper, we investigate the fine-grained clustering task and propose a novel model reuse framework. This framework outperforms existing fine-grained clustering methods by enhancing clustering consistency and robustness. For consistency, it employs low-rank optimization, which enforces stable and high-confidence predictions across augmented views of the same sample. For robustness, it leverages sparsification guided by reused models; this facilitates better handling of intra-class variances and inter-class similarities without converging to trivial solutions. We unify our model in a low rank optimization model. Specially, our model is guided by high-confidence groups through a reused model to perform sparsification of augmented matrices of different perturbations of the same sample to achieve low rank, thereby producing consistent and high-confidence clustering results. And we theoretically prove that low rank of sample augmented matrices can be achieved under our sparsification conditions, thus providing a powerful fine-grained unsupervised alternative. Our method achieves state-of-the-art clustering performance on three fine-grained image datasets.

深度聚类将深度模型扩展到聚类任务中，以其优异的聚类性能而备受关注。然而，传统的深度聚类方法假设不同类别的样本略有不同。当观察结果来自高度相似的样本时，实际应用是复杂的。在本文中，我们研究了这种细粒度聚类任务，而传统的粗粒度聚类难以捕捉细微的语义差异，这往往导致具有相似特征的聚类之间的决策边界不明确。我们的目标是学习特征表示，以鼓励细粒度数据在嵌入空间中形成清晰的聚类边界。本文研究了细粒度聚类任务，提出了一种新的模型重用框架。该框架通过增强聚类一致性和鲁棒性，优于现有的细粒度聚类方法。为了保持一致性，它采用低秩优化，在相同样本的增强视图中强制进行稳定和高置信度的预测。对于鲁棒性，它利用了由重用模型指导的稀疏化；这有助于更好地处理类内差异和类间相似性，而不会收敛到琐碎的解决方案。我们将模型统一为一个低阶优化模型。特别地，我们的模型通过重用模型以高置信度组为指导，对同一样本的不同扰动增强矩阵进行稀疏化，达到低秩，从而得到一致的高置信度聚类结果。我们从理论上证明了在我们的稀疏化条件下可以实现低秩的样本增广矩阵，从而提供了一个强大的细粒度无监督替代方案。我们的方法在三个细粒度图像数据集上实现了最先进的聚类性能。

{"title":"Deep fine-grained clustering with model reusing","authors":"Jie Hong, Xulun Ye, Jieyu Zhao","doi":"10.1016/j.neunet.2026.108604","DOIUrl":"10.1016/j.neunet.2026.108604","url":null,"abstract":"<div><div>Deep clustering, which extends deep models to the clustering task, has attracted many attentions due to the high clustering performance. However, conventional deep clustering method assume that sample in different class are slightly different. Real application is complicated where observations are achieved from highly similar samples. In this paper, we study this fine-grained clustering task, while traditional coarse-grained clustering has difficulty capturing subtle semantic differences, which often leads to unclear decision boundaries between clusters of similar features. Our goal is to learn feature representations that encourage fine-grained data to form clear cluster boundaries in the embedding space. In this paper, we investigate the fine-grained clustering task and propose a novel model reuse framework. This framework outperforms existing fine-grained clustering methods by enhancing clustering consistency and robustness. For consistency, it employs low-rank optimization, which enforces stable and high-confidence predictions across augmented views of the same sample. For robustness, it leverages sparsification guided by reused models; this facilitates better handling of intra-class variances and inter-class similarities without converging to trivial solutions. We unify our model in a low rank optimization model. Specially, our model is guided by high-confidence groups through a reused model to perform sparsification of augmented matrices of different perturbations of the same sample to achieve low rank, thereby producing consistent and high-confidence clustering results. And we theoretically prove that low rank of sample augmented matrices can be achieved under our sparsification conditions, thus providing a powerful fine-grained unsupervised alternative. Our method achieves state-of-the-art clustering performance on three fine-grained image datasets.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"198 ","pages":"Article 108604"},"PeriodicalIF":6.3,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145999310","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

Adaptive dynamic spatial-temporal graph convolutional neural network for traffic flow prediction 交通流预测的自适应动态时空图卷积神经网络

IF 6.3 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neural Networks

Pub Date : 2026-06-01 Epub Date: 2026-01-04 DOI: 10.1016/j.neunet.2025.108529

Yu Jiang , Mingmao Hu , Aihong Gong , Yanfei Lan , Qingshan Gong , Shuaiyu Li , Xu Wang , Zhenghao Yao

Accurate and efficient traffic flow prediction is essential for developing smart cities. Traffic flow data exhibits complex spatio-temporal dependencies, and the weights between nodes may change dynamically due to travel patterns and node attributes. However, existing prediction models primarily rely on static adjacency matrices and complex time series models, which limit model performance. We propose an innovative traffic flow prediction method, the Adaptive Dynamic Spatio-temporal Graph Convolutional Network (ADSTGCN), to address this issue. Specifically, we incorporate a multi-head attention mechanism and an adaptive dynamic adjacency matrix to construct two dynamic spatio-temporal extraction modules, which are integrated with Graph Convolutional Networks (GCN) to overcome the limitation of static adjacency matrices in capturing dynamic spatio-temporal correlations between nodes. At the temporal dimension extraction level, we integrate the Mamba to model long-term time series in traffic flow data, effectively extracting relevant temporal information. Extensive comparative experiments are conducted on four real-world public transportation datasets. The results demonstrate that our model achieves the highest prediction accuracy compared to other baseline models, showcasing its significant potential in traffic flow prediction.

准确、高效的交通流量预测对智慧城市的发展至关重要。交通流数据表现出复杂的时空依赖关系，节点间的权重会因出行模式和节点属性的不同而发生动态变化。然而，现有的预测模型主要依赖于静态邻接矩阵和复杂时间序列模型，这限制了模型的性能。为了解决这一问题，我们提出了一种创新的交通流预测方法——自适应动态时空图卷积网络（ADSTGCN）。具体而言，我们结合多头注意机制和自适应动态邻接矩阵构建了两个动态时空提取模块，并将其与图卷积网络（GCN）相结合，克服了静态邻接矩阵在捕获节点间动态时空相关性方面的局限性。在时间维度提取层面，我们将曼巴模型整合到交通流数据的长期时间序列中，有效提取相关的时间信息。在四个真实的公共交通数据集上进行了广泛的比较实验。结果表明，与其他基线模型相比，我们的模型达到了最高的预测精度，显示了其在交通流预测方面的巨大潜力。

{"title":"Adaptive dynamic spatial-temporal graph convolutional neural network for traffic flow prediction","authors":"Yu Jiang , Mingmao Hu , Aihong Gong , Yanfei Lan , Qingshan Gong , Shuaiyu Li , Xu Wang , Zhenghao Yao","doi":"10.1016/j.neunet.2025.108529","DOIUrl":"10.1016/j.neunet.2025.108529","url":null,"abstract":"<div><div>Accurate and efficient traffic flow prediction is essential for developing smart cities. Traffic flow data exhibits complex spatio-temporal dependencies, and the weights between nodes may change dynamically due to travel patterns and node attributes. However, existing prediction models primarily rely on static adjacency matrices and complex time series models, which limit model performance. We propose an innovative traffic flow prediction method, the Adaptive Dynamic Spatio-temporal Graph Convolutional Network (ADSTGCN), to address this issue. Specifically, we incorporate a multi-head attention mechanism and an adaptive dynamic adjacency matrix to construct two dynamic spatio-temporal extraction modules, which are integrated with Graph Convolutional Networks (GCN) to overcome the limitation of static adjacency matrices in capturing dynamic spatio-temporal correlations between nodes. At the temporal dimension extraction level, we integrate the Mamba to model long-term time series in traffic flow data, effectively extracting relevant temporal information. Extensive comparative experiments are conducted on four real-world public transportation datasets. The results demonstrate that our model achieves the highest prediction accuracy compared to other baseline models, showcasing its significant potential in traffic flow prediction.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"198 ","pages":"Article 108529"},"PeriodicalIF":6.3,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145980478","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

Temporal knowledge graphs forecasting based on explainable temporal relation tree-graph 基于可解释时间关系树形图的时间知识图预测

IF 6.3 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neural Networks

Pub Date : 2026-06-01 Epub Date: 2026-01-07 DOI: 10.1016/j.neunet.2026.108582

Qihong Wu , Ruizhe Ma , Yuan Cheng , Li Yan , Zongmin Ma

In real-world temporal knowledge graphs, relationships among entities often exhibit complex temporal dynamics. Effectively modeling multi-hop temporal relation chains and enabling interpretable reasoning remain core challenges in temporal knowledge graph forecasting, which we address with our proposed model, TRTL (Temporal Relation Tree-based Learning). To tackle these challenges, we introduce a novel reasoning framework grounded in two complementary graph structures: the Sequence Grounding Graph, which captures temporal interval and entity’s relation alignments; the Temporal Relation Tree Graph, which organizes multi-hop relation chains into interpretable and tree-structured reasoning paths. These structures are encoded using a Tree-LSTM enhanced with attention mechanisms, enabling the model to effectively capture temporal logic and long-range dependencies. The tree-based symbolical reasoning process provides interpretable evidence, enhancing the transparency and reliability of predictions. Experiments on two time-interval benchmarks demonstrate that TRTL significantly outperforms existing symbolic-based models.

在现实世界的时间知识图中，实体之间的关系往往表现出复杂的时间动态。有效地建模多跳时间关系链并实现可解释推理仍然是时间知识图预测的核心挑战，我们提出了基于时间关系树的学习模型TRTL （temporal relation Tree-based Learning）。为了解决这些挑战，我们引入了一种基于两个互补图结构的新型推理框架：序列接地图，它捕获时间间隔和实体的关系对齐；时间关系树图，它将多跳关系链组织成可解释的树结构推理路径。这些结构使用带有注意机制的Tree-LSTM进行编码，使模型能够有效地捕获时间逻辑和远程依赖关系。基于树的符号推理过程提供了可解释的证据，提高了预测的透明度和可靠性。在两个时间间隔基准上的实验表明，TRTL显著优于现有的基于符号的模型。

{"title":"Temporal knowledge graphs forecasting based on explainable temporal relation tree-graph","authors":"Qihong Wu , Ruizhe Ma , Yuan Cheng , Li Yan , Zongmin Ma","doi":"10.1016/j.neunet.2026.108582","DOIUrl":"10.1016/j.neunet.2026.108582","url":null,"abstract":"<div><div>In real-world temporal knowledge graphs, relationships among entities often exhibit complex temporal dynamics. Effectively modeling multi-hop temporal relation chains and enabling interpretable reasoning remain core challenges in temporal knowledge graph forecasting, which we address with our proposed model, TRTL (Temporal Relation Tree-based Learning). To tackle these challenges, we introduce a novel reasoning framework grounded in two complementary graph structures: the Sequence Grounding Graph, which captures temporal interval and entity’s relation alignments; the Temporal Relation Tree Graph, which organizes multi-hop relation chains into interpretable and tree-structured reasoning paths. These structures are encoded using a Tree-LSTM enhanced with attention mechanisms, enabling the model to effectively capture temporal logic and long-range dependencies. The tree-based symbolical reasoning process provides interpretable evidence, enhancing the transparency and reliability of predictions. Experiments on two time-interval benchmarks demonstrate that TRTL significantly outperforms existing symbolic-based models.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"198 ","pages":"Article 108582"},"PeriodicalIF":6.3,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145981713","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