Journal of Computational Science最新文献

GAT based attention-reweighted PPI networks uncover shared hub genes in Crohn’s disease and ulcerative colitis 基于GAT的注意力重加权PPI网络揭示了克罗恩病和溃疡性结肠炎的共享中心基因

IF 3.7 3区计算机科学 Q2 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Journal of Computational Science

Pub Date : 2026-01-28 DOI: 10.1016/j.jocs.2026.102791

Pratibha Joshi , Rinki Basoya , Rana Pratap Singh , Ravi Verma , Buddha Singh

Inflammatory bowel disease (IBD), comprising Crohn’s disease (CD) and ulcerative colitis (UC), exhibits molecular heterogeneity that complicates diagnosis. This study proposes an integrated deep learning and bioinformatics framework to uncover molecular pathways and identify potential therapeutic targets. The dataset includes 156 CD patients, 167 UC patients and 267 healthy controls. The Graph Attention Network (GAT) is applied to protein–protein interaction (PPI) networks for embedding the node features and reweighting edges based on learned attention scores. The Louvain algorithm on the embedded PPI network identifies the top 10 communities. Functional enrichment analysis of the genes within these communities reveals significantly enriched in immune responses, stress responses and pathogen-associated pathways. From these communities, 32 hub genes are identified as being implicated in IBD pathogenesis. The machine learning classifiers logistic regression, support vector machine, random forest, gradient boosting and a stacking classifier are applied to 32 genes to distinguish between CD and UC. Based on classification performance and statistical significance, 10 highly significant genes TLR5, TLR2, IL1B, IL4R, TLR4, IL18, STAT3, STAT1, IL18RAP and IFNGR1 are selected. Furthermore, the KEGG pathway analysis of the 32 genes shows that five of these genes (STAT1, STAT3, IL4R, IL18 and TLR2) are directly involved in IBD-related pathways. Experimental validation of five key genes using qRT-PCR in HCT116 cells confirms significant upregulation of these genes. Modularity and NMI scores demonstrate that the GAT-based framework achieves superior community detection performance compared to baseline methods. This approach gives a scalable and robust strategy for advancing biomarker discovery and personalized therapy in complex diseases like IBD.

炎症性肠病（IBD），包括克罗恩病（CD）和溃疡性结肠炎（UC），表现出分子异质性，使诊断复杂化。本研究提出了一个集成的深度学习和生物信息学框架，以揭示分子途径并确定潜在的治疗靶点。该数据集包括156名乳糜泻患者，167名UC患者和267名健康对照。将图注意网络（GAT）应用于蛋白质-蛋白质相互作用（PPI）网络中，用于嵌入节点特征并根据学习到的注意分数重新加权边缘。嵌入式PPI网络上的Louvain算法识别出排名前10位的社区。功能富集分析显示，这些群落中基因在免疫反应、应激反应和病原体相关途径中显著富集。从这些群体中，鉴定出32个中枢基因与IBD发病机制有关。将机器学习分类器逻辑回归、支持向量机、随机森林、梯度增强和堆叠分类器应用于32个基因，以区分CD和UC。根据分类性能和统计学显著性，选择10个高度显著基因TLR5、TLR2、IL1B、IL4R、TLR4、IL18、STAT3、STAT1、IL18RAP和IFNGR1。此外，对32个基因的KEGG通路分析显示，其中5个基因（STAT1、STAT3、IL4R、IL18和TLR2）直接参与ibd相关通路。在HCT116细胞中使用qRT-PCR对5个关键基因进行实验验证，证实了这些基因的显著上调。模块化和NMI分数表明，与基线方法相比，基于gat的框架具有优越的社区检测性能。这种方法为推进IBD等复杂疾病的生物标志物发现和个性化治疗提供了可扩展和强大的策略。

{"title":"GAT based attention-reweighted PPI networks uncover shared hub genes in Crohn’s disease and ulcerative colitis","authors":"Pratibha Joshi , Rinki Basoya , Rana Pratap Singh , Ravi Verma , Buddha Singh","doi":"10.1016/j.jocs.2026.102791","DOIUrl":"10.1016/j.jocs.2026.102791","url":null,"abstract":"<div><div>Inflammatory bowel disease (IBD), comprising Crohn’s disease (CD) and ulcerative colitis (UC), exhibits molecular heterogeneity that complicates diagnosis. This study proposes an integrated deep learning and bioinformatics framework to uncover molecular pathways and identify potential therapeutic targets. The dataset includes 156 CD patients, 167 UC patients and 267 healthy controls. The Graph Attention Network (GAT) is applied to protein–protein interaction (PPI) networks for embedding the node features and reweighting edges based on learned attention scores. The <strong>Louvain algorithm</strong> on the embedded PPI network identifies the top 10 communities. Functional enrichment analysis of the genes within these communities reveals significantly enriched in immune responses, stress responses and pathogen-associated pathways. From these communities, 32 hub genes are identified as being implicated in IBD pathogenesis. The <strong>machine learning classifiers</strong> logistic regression, support vector machine, random forest, gradient boosting and a stacking classifier are applied to 32 genes to distinguish between CD and UC. Based on classification performance and statistical significance, 10 highly significant genes <strong>TLR5, TLR2, IL1B, IL4R, TLR4, IL18, STAT3, STAT1, IL18RAP</strong> and <strong>IFNGR1</strong> are selected. Furthermore, the KEGG pathway analysis of the 32 genes shows that five of these genes (<strong>STAT1</strong>, <strong>STAT3</strong>, <strong>IL4R</strong>, <strong>IL18</strong> and <strong>TLR2)</strong> are directly involved in IBD-related pathways. Experimental validation of five key genes using qRT-PCR in HCT116 cells confirms significant upregulation of these genes. Modularity and NMI scores demonstrate that the GAT-based framework achieves superior community detection performance compared to baseline methods. This approach gives a scalable and robust strategy for advancing biomarker discovery and personalized therapy in complex diseases like IBD.</div></div>","PeriodicalId":48907,"journal":{"name":"Journal of Computational Science","volume":"95 ","pages":"Article 102791"},"PeriodicalIF":3.7,"publicationDate":"2026-01-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146081490","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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

Pattern embedding driven lightweight quadrilateral mesh generation for engineering applications 面向工程应用的模式嵌入驱动轻量级四边形网格生成

IF 3.7 3区计算机科学 Q2 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Journal of Computational Science

Pub Date : 2026-01-27 DOI: 10.1016/j.jocs.2026.102790

Junxian Liu , Minglin Li , Wei Ye , Jinxu Liu , Lianfeng Lai

Current mainstream engineering simulation software struggles to generate high-quality quadrilateral meshes during the preprocessing stage, and achieving such meshes typically relies on numerical solvers. To enable efficient and high-quality quadrilateral mesh generation for complex geometries in open-source environments, this paper proposes a lightweight quadrilateral mesh reconstruction method. The method first takes a triangular mesh model as input and parameterizes it onto a 2D plane and performs polygonal layout decomposition using directional constraints. A greedy propagation strategy is then employed to assign uniform subdivision parameters to layout boundaries, replacing global integer programming. Finally, quadrilateral mesh templates are embedded to complete the mesh generation. The entire process avoids reliance on numerical solvers. Experiments on publicly available benchmark models demonstrate that, even without using solvers, the generated meshes achieve quality comparable to or better than mainstream approaches, providing a low-cost, high-accuracy preprocessing solution for engineering simulations.

目前主流的工程仿真软件在预处理阶段难以生成高质量的四边形网格，而实现这种网格通常依赖于数值求解器。为了在开源环境下对复杂几何图形进行高效、高质量的四边形网格生成，提出了一种轻量级的四边形网格重建方法。该方法首先以三角形网格模型为输入，将其参数化到二维平面上，利用方向约束进行多边形布局分解。然后采用贪婪传播策略为布局边界分配统一的细分参数，取代全局整数规划。最后，嵌入四边形网格模板，完成网格生成。整个过程避免了对数值求解器的依赖。在公开可用的基准模型上进行的实验表明，即使不使用求解器，生成的网格也可以达到与主流方法相当或更好的质量，为工程仿真提供了低成本，高精度的预处理解决方案。

{"title":"Pattern embedding driven lightweight quadrilateral mesh generation for engineering applications","authors":"Junxian Liu , Minglin Li , Wei Ye , Jinxu Liu , Lianfeng Lai","doi":"10.1016/j.jocs.2026.102790","DOIUrl":"10.1016/j.jocs.2026.102790","url":null,"abstract":"<div><div>Current mainstream engineering simulation software struggles to generate high-quality quadrilateral meshes during the preprocessing stage, and achieving such meshes typically relies on numerical solvers. To enable efficient and high-quality quadrilateral mesh generation for complex geometries in open-source environments, this paper proposes a lightweight quadrilateral mesh reconstruction method. The method first takes a triangular mesh model as input and parameterizes it onto a 2D plane and performs polygonal layout decomposition using directional constraints. A greedy propagation strategy is then employed to assign uniform subdivision parameters to layout boundaries, replacing global integer programming. Finally, quadrilateral mesh templates are embedded to complete the mesh generation. The entire process avoids reliance on numerical solvers. Experiments on publicly available benchmark models demonstrate that, even without using solvers, the generated meshes achieve quality comparable to or better than mainstream approaches, providing a low-cost, high-accuracy preprocessing solution for engineering simulations.</div></div>","PeriodicalId":48907,"journal":{"name":"Journal of Computational Science","volume":"95 ","pages":"Article 102790"},"PeriodicalIF":3.7,"publicationDate":"2026-01-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146057291","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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

Exploring scalarization methods and approximation algorithms for the Multi-Objective Set Covering Problem 探索多目标集覆盖问题的标量化方法和逼近算法

IF 3.7 3区计算机科学 Q2 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Journal of Computational Science

Pub Date : 2026-01-23 DOI: 10.1016/j.jocs.2026.102785

Lakmali Weerasena, Chathuri Aththanayake

This study focuses on the Multi-Objective Set Covering Problem (MOSCP), a prominent challenge in Multi-Objective Combinatorial Optimization. Given the NP-completeness of the Set Covering Problem (SCP), both Single Objective SCP (SOSCP) and MOSCP evolve into NP-hard combinatorial optimization problems. Our investigation reveals a gap in the literature, as norm-based scalarizations have been identified as effective in approximating multi-objective minimization problems, yet limited research exists on

ϵ

-approximation algorithms for the MOSCP. Our contribution unfolds on two fronts: (1) We explore the impact of using a common set of weight vectors for two scalarization methods weighted-sum and weighted max-ordering in solving the MOSCP under various conditions. We integrate four distinct weight computing methods and precisely solve both weighted-sum scalarization and weighted max-ordering scalarization of the MOSCP. (2) We utilize the cost-effectiveness vector proposed in previous literature to extract additional theoretical properties describing the approximation quality. Furthermore, we introduce a new cost-effectiveness vector and discuss its validity in relation to the properties identified in prior studies, resulting in the derivation of new theoretical properties. The study indicates that introduced weight generation methods demonstrate commendable performance with an equivalent number of weight vectors on the MOSCP. Furthermore, employing weighted max-ordering scalarization is proposed as a suitable choice for generating a robust initial pool for two-phase algorithms.

研究了多目标集合覆盖问题，这是多目标组合优化中的一个突出问题。给定集合覆盖问题（SCP）的np完备性，单目标覆盖问题（SOSCP）和多目标覆盖问题都演化为np困难组合优化问题。我们的研究揭示了文献中的一个空白，因为基于范数的标量化已经被认为是近似多目标最小化问题的有效方法，但关于ϵ-approximation算法的MOSCP研究有限。我们的贡献体现在两个方面：(1)我们探索了在不同条件下，使用一组共同的权向量对加权和和和加权最大排序两种标量化方法在求解MOSCP问题中的影响。我们整合了四种不同的权重计算方法，精确地解决了MOSCP的加权和标量化和加权最大阶标量化问题。(2)我们利用先前文献中提出的成本效益向量来提取描述近似质量的附加理论性质。此外，我们引入了一个新的成本效益向量，并讨论了其与先前研究中确定的性质相关的有效性，从而推导出新的理论性质。研究表明，所引入的权值生成方法在MOSCP上具有相当数量的权值向量，具有良好的性能。在此基础上，提出了采用加权最大阶标量化来生成两阶段算法的鲁棒初始池的合适选择。

{"title":"Exploring scalarization methods and approximation algorithms for the Multi-Objective Set Covering Problem","authors":"Lakmali Weerasena, Chathuri Aththanayake","doi":"10.1016/j.jocs.2026.102785","DOIUrl":"10.1016/j.jocs.2026.102785","url":null,"abstract":"<div><div>This study focuses on the Multi-Objective Set Covering Problem (MOSCP), a prominent challenge in Multi-Objective Combinatorial Optimization. Given the NP-completeness of the Set Covering Problem (SCP), both Single Objective SCP (SOSCP) and MOSCP evolve into NP-hard combinatorial optimization problems. Our investigation reveals a gap in the literature, as norm-based scalarizations have been identified as effective in approximating multi-objective minimization problems, yet limited research exists on <span><math><mi>ϵ</mi></math></span>-approximation algorithms for the MOSCP. Our contribution unfolds on two fronts: (1) We explore the impact of using a common set of weight vectors for two scalarization methods weighted-sum and weighted max-ordering in solving the MOSCP under various conditions. We integrate four distinct weight computing methods and precisely solve both weighted-sum scalarization and weighted max-ordering scalarization of the MOSCP. (2) We utilize the cost-effectiveness vector proposed in previous literature to extract additional theoretical properties describing the approximation quality. Furthermore, we introduce a new cost-effectiveness vector and discuss its validity in relation to the properties identified in prior studies, resulting in the derivation of new theoretical properties. The study indicates that introduced weight generation methods demonstrate commendable performance with an equivalent number of weight vectors on the MOSCP. Furthermore, employing weighted max-ordering scalarization is proposed as a suitable choice for generating a robust initial pool for two-phase algorithms.</div></div>","PeriodicalId":48907,"journal":{"name":"Journal of Computational Science","volume":"94 ","pages":"Article 102785"},"PeriodicalIF":3.7,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146022547","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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

Numerical method-informed DeepONet for refractivity inversion in waveguides 基于数值方法的波导折射率反演DeepONet

IF 3.7 3区计算机科学 Q2 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Journal of Computational Science

Pub Date : 2026-01-23 DOI: 10.1016/j.jocs.2026.102788

Mikhail S. Lytaev

This work applies deep learning methods to estimate vertical refractive index profiles in elongated waveguides. We use the DeepONet architecture to learn an inverse operator that maps signal measurements from a known source to the refractive index profile. The forward model is the one-way Helmholtz equation. A variational autoencoder is employed to augment the input data used for training the inverse operator. The obtained solution is then refined using the automatically differentiable forward model. Computational experiments are performed for tropospheric and underwater tomography problems, including experiments on real data. The numerical results confirm the effectiveness of the proposed approach. A Python 3 (JAX) implementation of the proposed method is publicly available. This work is an extended version of the ICCS-2025 conference paper (Lytaev, 2025).

这项工作应用深度学习方法来估计细长波导的垂直折射率分布。我们使用DeepONet架构来学习一个逆算子，该算子将来自已知源的信号测量映射到折射率剖面。正向模型是单向亥姆霍兹方程。采用变分自编码器增强用于训练逆算子的输入数据。然后使用自动可微正演模型对得到的解进行细化。对对流层和水下层析成像问题进行了计算实验，包括在实际数据上的实验。数值结果验证了该方法的有效性。该方法的Python 3 （JAX）实现是公开的。这项工作是ICCS-2025会议论文（Lytaev, 2025）的扩展版本。

引用次数: 0

Single/multi step optimal and modified homotopy perturbation method for strongly non-linear fractional initial value problems: Global series solution 强非线性分数阶初值问题的单/多步最优修正同伦摄动方法：全局级数解

IF 3.7 3区计算机科学 Q2 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Journal of Computational Science

Pub Date : 2026-01-16 DOI: 10.1016/j.jocs.2026.102786

Tapas Roy, Dilip K. Maiti

The aim of this work is to explore the possibility for obtaining globally convergent series solutions for fractional order initial value problems (FIVPs) using our recently developed semi-analytical technique, known as the optimal and modified homotopy perturbation method (OM-HPM). First time here we demonstrate the phenomenon of global series solutions for both IVPs and FIVPs, providing guidelines for selecting the linear operator and initial guess when applying homotopy methods to fractional differential equations. One of our noble aims is to establish the necessary and sufficient conditions for the convergence of solutions for all contemporary homotopy-based methods, including OM-HPM, and to validate these conditions numerically. To illustrate the accuracy and efficiency of our technique, we apply it to six strongly nonlinear steady as well as chaotic fractional order IVPs, comparing our results with exact solutions and those obtained using numerical methods and other contemporary semi-analytical approaches. We exemplify the broader applicability of our single-step OM-HPM method in achieving global solutions for FIVPs, in contrast to contemporary homotopy-based multi-step methods. Additionally, we investigate the limitations of OM-HPM and introduce the multi-step OM-HPM, or MOM-HPM, specifically designed for chaotic solutions of the Lorenz system. We also provide recommendations on selecting the appropriate step size for all previously mentioned multi-step homotopy methods, aiming to minimize computational time within the designated domain. Our comprehensive theoretical and numerical results demonstrate that the single-step OM-HPM method is more advanced and effective than other existing single and multi-step methods for obtaining global series solutions for strongly nonlinear FIVPs.

这项工作的目的是探索利用我们最近开发的半解析技术，即最优和改进同伦摄动法（OM-HPM），获得分数阶初值问题（FIVPs）全局收敛级数解的可能性。本文首次展示了ivp和fivp的全局级数解现象，为分数阶微分方程同伦方法的线性算子和初始猜想的选择提供了指导。我们的崇高目标之一是建立所有当代基于同伦的方法（包括OM-HPM）解收敛的充分必要条件，并对这些条件进行数值验证。为了说明我们的技术的准确性和效率，我们将其应用于六个强非线性稳态和混沌分数阶ivp，并将我们的结果与精确解以及使用数值方法和其他当代半解析方法获得的结果进行比较。与当代基于同伦的多步骤方法相比，我们的单步OM-HPM方法在实现fivp全局解决方案方面具有更广泛的适用性。此外，我们研究了OM-HPM的局限性，并介绍了专门设计用于洛伦兹系统混沌解的多步OM-HPM或MOM-HPM。我们还为前面提到的所有多步同伦方法选择合适的步长提供了建议，旨在减少指定域内的计算时间。综合理论和数值结果表明，单步OM-HPM方法在求解强非线性fivp全局级数解方面比现有的单步和多步方法更为先进和有效。

{"title":"Single/multi step optimal and modified homotopy perturbation method for strongly non-linear fractional initial value problems: Global series solution","authors":"Tapas Roy, Dilip K. Maiti","doi":"10.1016/j.jocs.2026.102786","DOIUrl":"10.1016/j.jocs.2026.102786","url":null,"abstract":"<div><div>The aim of this work is to explore the possibility for obtaining globally convergent series solutions for fractional order initial value problems (FIVPs) using our recently developed semi-analytical technique, known as the optimal and modified homotopy perturbation method (OM-HPM). First time here we demonstrate the phenomenon of global series solutions for both IVPs and FIVPs, providing guidelines for selecting the linear operator and initial guess when applying homotopy methods to fractional differential equations. One of our noble aims is to establish the necessary and sufficient conditions for the convergence of solutions for all contemporary homotopy-based methods, including OM-HPM, and to validate these conditions numerically. To illustrate the accuracy and efficiency of our technique, we apply it to six strongly nonlinear steady as well as chaotic fractional order IVPs, comparing our results with exact solutions and those obtained using numerical methods and other contemporary semi-analytical approaches. We exemplify the broader applicability of our single-step OM-HPM method in achieving global solutions for FIVPs, in contrast to contemporary homotopy-based multi-step methods. Additionally, we investigate the limitations of OM-HPM and introduce the multi-step OM-HPM, or MOM-HPM, specifically designed for chaotic solutions of the Lorenz system. We also provide recommendations on selecting the appropriate step size for all previously mentioned multi-step homotopy methods, aiming to minimize computational time within the designated domain. Our comprehensive theoretical and numerical results demonstrate that the single-step OM-HPM method is more advanced and effective than other existing single and multi-step methods for obtaining global series solutions for strongly nonlinear FIVPs.</div></div>","PeriodicalId":48907,"journal":{"name":"Journal of Computational Science","volume":"94 ","pages":"Article 102786"},"PeriodicalIF":3.7,"publicationDate":"2026-01-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146022641","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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 joint graph neural network model incorporating rhetorical structure theory 结合修辞结构理论的联合图神经网络模型

IF 3.7 3区计算机科学 Q2 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Journal of Computational Science

Pub Date : 2026-01-14 DOI: 10.1016/j.jocs.2026.102784

Xiaoyang Wang , Wenfeng Liu , Xuesong Jiang , Jiangwei Wang , Yuzhen Yang , Yaling Gao , Longqing Bao

The rapid development of graph neural networks (GNNs) has led to significant advances in text classification. Current approaches primarily focus on converting text into graph representations by modeling word relationships, achieving promising results in natural language processing tasks. However, these methods often overlook crucial discourse-level information and the hierarchical organization of text documents. This paper introduces a novel framework that leverages Rhetorical Structure Theory (RST) to capture document-level discourse structure and proposes a multi-graph joint learning approach. Our main contributions are: (1) we propose the first framework to systematically integrate RST-based discourse structure with word-level features for neural text classification, (2) we develop methods to construct RST graphs that effectively preserve hierarchical discourse information, and (3) we design RSTGNN, a multi-graph joint learning architecture that combines discourse structure with semantic, syntactic, and sequential information through specialized attention mechanisms. Extensive experiments on five text classification datasets demonstrate that our approach achieves competitive performance with notable improvements on several benchmark datasets.

图神经网络（gnn）的快速发展使文本分类技术取得了重大进展。目前的方法主要集中在通过建模词关系将文本转换为图形表示，在自然语言处理任务中取得了有希望的结果。然而，这些方法往往忽略了关键的话语级信息和文本文档的层次组织。本文介绍了一种利用修辞结构理论（RST）捕捉文档级语篇结构的新框架，并提出了一种多图联合学习方法。我们的主要贡献有：(1)我们提出了第一个将基于RST的话语结构与词级特征系统集成的框架，用于神经文本分类；(2)我们开发了构建RST图的方法，有效地保留了分层话语信息；(3)我们设计了RSTGNN，这是一种多图联合学习架构，通过专门的注意机制将话语结构与语义、句法和顺序信息相结合。在五个文本分类数据集上的大量实验表明，我们的方法在几个基准数据集上取得了显著的改进，取得了具有竞争力的性能。

{"title":"A joint graph neural network model incorporating rhetorical structure theory","authors":"Xiaoyang Wang , Wenfeng Liu , Xuesong Jiang , Jiangwei Wang , Yuzhen Yang , Yaling Gao , Longqing Bao","doi":"10.1016/j.jocs.2026.102784","DOIUrl":"10.1016/j.jocs.2026.102784","url":null,"abstract":"<div><div>The rapid development of graph neural networks (GNNs) has led to significant advances in text classification. Current approaches primarily focus on converting text into graph representations by modeling word relationships, achieving promising results in natural language processing tasks. However, these methods often overlook crucial discourse-level information and the hierarchical organization of text documents. This paper introduces a novel framework that leverages Rhetorical Structure Theory (RST) to capture document-level discourse structure and proposes a multi-graph joint learning approach. Our main contributions are: (1) we propose the first framework to systematically integrate RST-based discourse structure with word-level features for neural text classification, (2) we develop methods to construct RST graphs that effectively preserve hierarchical discourse information, and (3) we design RSTGNN, a multi-graph joint learning architecture that combines discourse structure with semantic, syntactic, and sequential information through specialized attention mechanisms. Extensive experiments on five text classification datasets demonstrate that our approach achieves competitive performance with notable improvements on several benchmark datasets.</div></div>","PeriodicalId":48907,"journal":{"name":"Journal of Computational Science","volume":"94 ","pages":"Article 102784"},"PeriodicalIF":3.7,"publicationDate":"2026-01-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146022545","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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

RBF partition of unity method for solving nonlinear multi-term fractional parabolic PDEs on irregular domains: A numerical study and error analysis 求解不规则区域上非线性多项分数抛物型偏微分方程的RBF分割统一方法：数值研究与误差分析

IF 3.7 3区计算机科学 Q2 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Journal of Computational Science

Pub Date : 2026-01-10 DOI: 10.1016/j.jocs.2026.102783

Banafsheh Raeisi, Mojtaba Fardi, Mohammadreza Ahmadi Darani

This paper introduces an innovative radial basis function partition of unity method for the numerical simulation of multi-term time-fractional parabolic partial differential equations. The proposed method simplifies algorithms and reduces computational costs by utilizing local approximations rather than global ones, thereby eliminating the need for differential operators on weight functions. It combines localized spatial discretization with a time-stepping strategy with a variable-step, which improves the rate of temporal convergence. The method uses the scalability of polyharmonic spline kernels to manage instability. The effectiveness of the proposed method is demonstrated through various numerical experiments.

本文介绍了一种新颖的径向基函数分划统一方法，用于多项时间分数型抛物型偏微分方程的数值模拟。该方法利用局部近似而非全局近似简化了算法，降低了计算成本，从而消除了对权函数的微分算子的需要。它将局部空间离散化与变步长时间步进策略相结合，提高了时间收敛速度。该方法利用多谐样条核的可扩展性来控制系统的不稳定性。通过各种数值实验验证了该方法的有效性。

引用次数: 0

On qualitative uncertainty in modelling assumptions 关于建模假设中的定性不确定性

IF 3.7 3区计算机科学 Q2 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Journal of Computational Science

Pub Date : 2026-01-02 DOI: 10.1016/j.jocs.2025.102782

Derek Groen, Laura M. Harbach

Researchers today have a range of advanced and efficient methods for quantifying uncertainty at their disposal. These methods effectively help them to understand how simulation results may change when a model is re-run or when input parameters are varied. However, models often contain assumptions that are not numerical or have uncertainties that cannot be quantified. Examples include assumed omissions, existing assumptions reused in new contexts, or assumptions based on partial evidence. This paper proposes a novel conceptual framework to investigate the uncertainty of modelling assumptions on a qualitative level. We aim to educate model developers on how to assess model quality beyond quantifiable uncertainties, understand how it can deteriorate, and identify measures that can improve quality or mitigate deterioration. The framework is designed to be broadly applicable to implemented models (simulations), conceptual models, and even mental models.

今天的研究人员有一系列先进而有效的方法来量化不确定性。这些方法有效地帮助他们了解当模型重新运行或输入参数变化时，模拟结果可能会发生什么变化。然而，模型通常包含非数值的假设或具有无法量化的不确定性。例子包括假设的遗漏，在新环境中重复使用的现有假设，或基于部分证据的假设。本文提出了一个新的概念框架，在定性水平上研究建模假设的不确定性。我们的目标是教育模型开发人员如何在可量化的不确定性之外评估模型质量，了解它是如何恶化的，并确定可以提高质量或减轻恶化的措施。该框架被设计成广泛适用于实现的模型（模拟）、概念模型，甚至心智模型。

引用次数: 0

Graph anomaly detection via local contrastive learning and attribute reconstruction 基于局部对比学习和属性重构的图异常检测

IF 3.7 3区计算机科学 Q2 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Journal of Computational Science

Pub Date : 2026-01-02 DOI: 10.1016/j.jocs.2025.102781

Xinyu Zhang , Yanfen Li

Graph anomaly detection (GAD) plays a vital role in identifying abnormal nodes within graph data, with applications in social networks, fraud detection, and cybersecurity. However, existing approaches face challenges such as anomaly overfitting and the homophily trap, which hinder accurate detection. Anomaly overfitting occurs when models are excessively influenced by anomalies during training, while the homophily trap arises from the assumption that connected nodes share similar features, which can mislead the model. To address these challenges, we propose Contrastive Local Anomaly-Aware Reconstruction Embedding (CLARE), a novel method designed to overcome these limitations. CLARE employs a contrastive sampling strategy to construct a local reference distribution, thus enhancing the learning of normal patterns. It integrates attribute and structural reconstruction to detect anomalies based on reconstruction errors. A key innovation is the neighbor and central node reconstruction mechanism, which improves detection accuracy by incorporating second-order neighbor information. Experimental results demonstrate that CLARE outperforms existing methods, offering robust and scalable anomaly detection for complex graph data.

图异常检测（GAD）在识别图数据中的异常节点方面起着至关重要的作用，在社交网络、欺诈检测和网络安全等领域都有应用。然而，现有的方法面临异常过拟合和同态陷阱等挑战，阻碍了准确的检测。当模型在训练过程中受到异常的过度影响时，就会出现异常过拟合，而假设连接节点具有相似的特征，就会产生同质陷阱，这可能会误导模型。为了解决这些问题，我们提出了一种新的方法——局部异常感知重构嵌入（CLARE），旨在克服这些局限性。CLARE采用对比采样策略构建局部参考分布，从而增强了对正态模式的学习。结合属性重构和结构重构，基于重构误差检测异常。其中一个关键的创新是邻居和中心节点重建机制，该机制通过纳入二阶邻居信息来提高检测精度。实验结果表明，该方法优于现有方法，对复杂图形数据提供了鲁棒性和可扩展性的异常检测。

{"title":"Graph anomaly detection via local contrastive learning and attribute reconstruction","authors":"Xinyu Zhang , Yanfen Li","doi":"10.1016/j.jocs.2025.102781","DOIUrl":"10.1016/j.jocs.2025.102781","url":null,"abstract":"<div><div>Graph anomaly detection (GAD) plays a vital role in identifying abnormal nodes within graph data, with applications in social networks, fraud detection, and cybersecurity. However, existing approaches face challenges such as anomaly overfitting and the homophily trap, which hinder accurate detection. Anomaly overfitting occurs when models are excessively influenced by anomalies during training, while the homophily trap arises from the assumption that connected nodes share similar features, which can mislead the model. To address these challenges, we propose Contrastive Local Anomaly-Aware Reconstruction Embedding (CLARE), a novel method designed to overcome these limitations. CLARE employs a contrastive sampling strategy to construct a local reference distribution, thus enhancing the learning of normal patterns. It integrates attribute and structural reconstruction to detect anomalies based on reconstruction errors. A key innovation is the neighbor and central node reconstruction mechanism, which improves detection accuracy by incorporating second-order neighbor information. Experimental results demonstrate that CLARE outperforms existing methods, offering robust and scalable anomaly detection for complex graph data.</div></div>","PeriodicalId":48907,"journal":{"name":"Journal of Computational Science","volume":"94 ","pages":"Article 102781"},"PeriodicalIF":3.7,"publicationDate":"2026-01-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146022642","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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 carbon neutral Chinese knowledge graph completion method based on latent semantic mining 基于潜在语义挖掘的时间碳中性中文知识图谱补全方法

IF 3.7 3区计算机科学 Q2 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Journal of Computational Science

Pub Date : 2025-12-31 DOI: 10.1016/j.jocs.2025.102744

Xiping Zhu, Lijuan Xiao, Ang Gao, Lu Guo, Huan Yang

In order to promote the digital transformation in the field of carbon neutrality, the article constructed a carbon neutral data set CANdata15k with a sample size of 16,050 to realize the completion of temporal knowledge graphs. The latent semantics of the data are not fully taken into account by the current temporal knowledge graph completion approach. In order to mine the entire latent semantics, we suggest a temporal carbon-neutral knowledge graph completion model (Tc-MLS). Replace entities and relationships in the pre-trained language model with text descriptions to acquire probable text semantics. And through the knowledge fusion model proposed in Tc-MLS to obtain the latent semantics between entities. The latent semantics between entities and relations are then obtained by embedding related topology and similarity soft logic. Finally, the comparison with other SOTA algorithm results shows that the MRR value of Tc-MLS on the CANdata15k dataset has increased by 56.91% on average, and the MRR values obtained on the two public datasets YAGO11k and Wikidata12k have increased by an average of 9.96% and 7.72%, respectively. Prove the effectiveness of the Tc-MLS model.

为了促进碳中和领域的数字化转型，本文构建了一个碳中和数据集CANdata15k，样本量为16050，实现了时间知识图的完成。目前的时态知识图补全方法没有充分考虑数据的潜在语义。为了挖掘整个潜在语义，我们提出了一个时间碳中性知识图补全模型（Tc-MLS）。用文本描述替换预训练语言模型中的实体和关系，以获得可能的文本语义。并通过在Tc-MLS中提出的知识融合模型获得实体之间的潜在语义。然后通过嵌入相关拓扑和相似软逻辑，获得实体和关系之间的潜在语义。最后，与其他SOTA算法结果的对比表明，Tc-MLS在CANdata15k数据集上的MRR值平均提高了56.91%，在YAGO11k和Wikidata12k两个公共数据集上的MRR值平均分别提高了9.96%和7.72%。验证了Tc-MLS模型的有效性。

{"title":"Temporal carbon neutral Chinese knowledge graph completion method based on latent semantic mining","authors":"Xiping Zhu, Lijuan Xiao, Ang Gao, Lu Guo, Huan Yang","doi":"10.1016/j.jocs.2025.102744","DOIUrl":"10.1016/j.jocs.2025.102744","url":null,"abstract":"<div><div>In order to promote the digital transformation in the field of carbon neutrality, the article constructed a carbon neutral data set CANdata15k with a sample size of 16,050 to realize the completion of temporal knowledge graphs. The latent semantics of the data are not fully taken into account by the current temporal knowledge graph completion approach. In order to mine the entire latent semantics, we suggest a temporal carbon-neutral knowledge graph completion model (Tc-MLS). Replace entities and relationships in the pre-trained language model with text descriptions to acquire probable text semantics. And through the knowledge fusion model proposed in Tc-MLS to obtain the latent semantics between entities. The latent semantics between entities and relations are then obtained by embedding related topology and similarity soft logic. Finally, the comparison with other SOTA algorithm results shows that the MRR value of Tc-MLS on the CANdata15k dataset has increased by 56.91% on average, and the MRR values obtained on the two public datasets YAGO11k and Wikidata12k have increased by an average of 9.96% and 7.72%, respectively. Prove the effectiveness of the Tc-MLS model.</div></div>","PeriodicalId":48907,"journal":{"name":"Journal of Computational Science","volume":"94 ","pages":"Article 102744"},"PeriodicalIF":3.7,"publicationDate":"2025-12-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145925533","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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