Applied Soft Computing最新文献

A discrete Keplerian optimization algorithm with Q-learning for human-robot collaborative partial disassembly line balancing problem 基于q -学习的离散Keplerian优化算法求解人机协作部分拆装线平衡问题

IF 6.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Applied Soft Computing

Pub Date : 2025-12-09 DOI: 10.1016/j.asoc.2025.114406

Zenan Tan , Shunsheng Guo , Jun Guo , Baigang Du , Kaipu Wang

Driven by advancements in human–machine collaboration technology, waste recycling enterprises are increasingly pursuing innovative production methodologies. Nevertheless, in practical human–machine collaborative partial disassembly tasks, task variability, process uncertainty, and multi-objective conflicts persist, rendering existing approaches inadequate for balancing efficiency, economic performance, and energy consumption. To address these challenges, this paper introduces a discrete Kepler optimization algorithm integrated with Q-learning (QLS-KOA) to resolve the human–machine collaborative disassembly line balancing problem (HRC-PDLBP), with the objective of simultaneously optimizing load distribution, disassembly profit, and energy consumption. At first, a two-dimensional coding scheme is designed to transform the continuous Kepler optimization algorithm into a discrete form, thereby generating high-quality initial solutions well-suited to the problem’s characteristics. Then, a Q-learning based local search framework is proposed to dynamically adjust the proportion of four HRC-PDLBP-specific local search operators in response to variations in the solution state. Considering that the solution space of HRC-PDLBP can dynamically change with task attributes and constraints, a Q-learning local search framework is introduced to make the intelligent agents efficiently capture optimization directions, discover global optimal solutions, and avoid local optima in dynamic environments. Finally, comparative experiments on multiple benchmark instances indicate that QLS-KOA achieves higher solution accuracy, faster convergence, and better solution diversity than state-of-the-art algorithms, while simultaneously reducing energy consumption and improving disassembly profit. Furthermore, a real case study on Tesla Model S battery module confirms that QLS-KOA generates practical and efficient disassembly schemes, and demonstrates significant advantages in terms of overall performance, sustainability, and industrial applicability.

在人机协作技术进步的推动下，废物回收企业越来越多地追求创新的生产方法。然而，在实际的人机协作部分拆卸任务中，任务可变性、过程不确定性和多目标冲突持续存在，使得现有方法不足以平衡效率、经济性能和能源消耗。为了解决这些挑战，本文引入了一种集成q -学习的离散Kepler优化算法（QLS-KOA）来解决人机协同拆解线平衡问题（HRC-PDLBP），目的是同时优化负载分配、拆卸利润和能耗。首先，设计了一种二维编码方案，将连续的Kepler优化算法转化为离散形式，从而生成适合问题特征的高质量初始解。然后，提出了一个基于q学习的局部搜索框架，根据解状态的变化动态调整4个hrc - pdlbp特定局部搜索算子的比例。考虑到HRC-PDLBP的解空间会随着任务属性和约束条件的变化而动态变化，引入q -学习局部搜索框架，使智能体在动态环境中高效捕获优化方向，发现全局最优解，避免局部最优。最后，在多个基准实例上的对比实验表明，与现有算法相比，QLS-KOA具有更高的解精度、更快的收敛速度和更好的解多样性，同时降低了能耗，提高了拆卸利润。此外，对特斯拉Model S电池模块的实际案例研究证实，QLS-KOA产生了实用高效的拆卸方案，在整体性能、可持续性和工业适用性方面具有显著优势。

{"title":"A discrete Keplerian optimization algorithm with Q-learning for human-robot collaborative partial disassembly line balancing problem","authors":"Zenan Tan , Shunsheng Guo , Jun Guo , Baigang Du , Kaipu Wang","doi":"10.1016/j.asoc.2025.114406","DOIUrl":"10.1016/j.asoc.2025.114406","url":null,"abstract":"<div><div>Driven by advancements in human–machine collaboration technology, waste recycling enterprises are increasingly pursuing innovative production methodologies. Nevertheless, in practical human–machine collaborative partial disassembly tasks, task variability, process uncertainty, and multi-objective conflicts persist, rendering existing approaches inadequate for balancing efficiency, economic performance, and energy consumption. To address these challenges, this paper introduces a discrete Kepler optimization algorithm integrated with Q-learning (QLS-KOA) to resolve the human–machine collaborative disassembly line balancing problem (HRC-PDLBP), with the objective of simultaneously optimizing load distribution, disassembly profit, and energy consumption. At first, a two-dimensional coding scheme is designed to transform the continuous Kepler optimization algorithm into a discrete form, thereby generating high-quality initial solutions well-suited to the problem’s characteristics. Then, a Q-learning based local search framework is proposed to dynamically adjust the proportion of four HRC-PDLBP-specific local search operators in response to variations in the solution state. Considering that the solution space of HRC-PDLBP can dynamically change with task attributes and constraints, a Q-learning local search framework is introduced to make the intelligent agents efficiently capture optimization directions, discover global optimal solutions, and avoid local optima in dynamic environments. Finally, comparative experiments on multiple benchmark instances indicate that QLS-KOA achieves higher solution accuracy, faster convergence, and better solution diversity than state-of-the-art algorithms, while simultaneously reducing energy consumption and improving disassembly profit. Furthermore, a real case study on Tesla Model S battery module confirms that QLS-KOA generates practical and efficient disassembly schemes, and demonstrates significant advantages in terms of overall performance, sustainability, and industrial applicability.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"188 ","pages":"Article 114406"},"PeriodicalIF":6.6,"publicationDate":"2025-12-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145749306","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 refined YOLOv5n-based method for detecting pepper flower objects integrating transfer learning 结合迁移学习的基于yolov5n的改进辣椒花物体检测方法

IF 6.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Applied Soft Computing

Pub Date : 2025-12-09 DOI: 10.1016/j.asoc.2025.114400

Minqiu Kuang , Fangping Xie , Dawei Liu , Bei Wu , Shang Chen , Yang Xiang , Feng Liu , Yanhua Wu , Xu Li

To address the low detection accuracy of small pepper flower targets in complex field environments and the need for extensive annotated datasets, a refined object detection model, YOLOv5n_Pepper Flower, based on an improved YOLOv5n architecture and transfer learning, is proposed in this study. First, to enhance detection accuracy and reduce the parameter count, the conventional convolution unit within the Neck layer of the YOLOv5n model was replaced with the simplified Grouped and Shuffled Convolution (GSConv). Second, the lightweight upsampling mechanism Content Aware Feature ReAssembly (CARAFE) was integrated into the Neck layer, significantly improving accuracy with minimal additional parameters. Third, the original PANet was replaced by the Bidirectional Feature Pyramid Network (BiFPN) structure, enhancing the fusion of multi-scale features through feature weighting. Lastly, the introduction of the attention mechanism BiFormer (Vision Transformer with Bi-Level Routing Attention) into the Neck layer improved the detection capability for small pepper flower targets, consequently boosting overall model performance. Experimental results showed that the refined YOLOv5n_Pepper Flower achieved a mean average precision (mAP) of 97 %, surpassing YOLOv5n, YOLOv5s, YOLOv7tiny, YOLOv8s, and YOLOv8n by margins of 0.9 %, 0.7 %, 5.9 %, 1.3 %, and 1.7 % respectively. Additionally, the comparison between the refined YOLOv5n_Pepper Flower and YOLOv9 indicated similar detection accuracies. However, the YOLOv5n-based model exhibits superiority in terms of computation, the number of parameters, and model size, and its lightweight model was superior regarding computational efficiency, parameter count, and model size. Its lightweight characteristics are more suitable for mobile terminal deployment. Furthermore, training the refined YOLOv5n model with transfer learning increased the mAP to 97.1 %, accelerated model convergence, and improved detection performance. The findings of this study provide a solid foundation for developing methods for pepper flower posture estimation and mechanical pollination.

针对复杂野外环境下小辣椒花目标检测精度低、需要大量标注数据集的问题，提出了一种基于改进YOLOv5n结构和迁移学习的改进目标检测模型YOLOv5n_Pepper flower。首先，为了提高检测精度和减少参数数量，将YOLOv5n模型的Neck层内的常规卷积单元替换为简化的分组和shuffle卷积（GSConv）。其次，将轻量级上采样机制Content Aware Feature ReAssembly （CARAFE）集成到Neck层中，以最少的额外参数显著提高了精度。第三，用双向特征金字塔网络（Bidirectional Feature Pyramid Network, BiFPN）结构取代原有的PANet，通过特征加权增强多尺度特征的融合。最后，在颈部层引入了注意力机制BiFormer (Vision Transformer with Bi-Level Routing attention)，提高了对小辣椒花目标的检测能力，从而提高了模型的整体性能。实验结果表明，改进后的YOLOv5n_Pepper Flower的平均精度（mAP）为97 %，分别比YOLOv5n、YOLOv5s、YOLOv7tiny、YOLOv8s和YOLOv8n高出0.9 %、0.7 %、5.9 %、1.3 %和1.7 %。此外，YOLOv5n_Pepper Flower与YOLOv9的检测精度相近。然而，基于yolov5n的模型在计算量、参数数量和模型尺寸方面具有优势，其轻量化模型在计算效率、参数数量和模型尺寸方面具有优势。其轻量级特性更适合移动终端部署。此外，使用迁移学习训练改进的YOLOv5n模型，将mAP提高到97.1 %，加速了模型收敛，提高了检测性能。研究结果为辣椒花位估算和机械授粉方法的开发提供了基础。

{"title":"A refined YOLOv5n-based method for detecting pepper flower objects integrating transfer learning","authors":"Minqiu Kuang , Fangping Xie , Dawei Liu , Bei Wu , Shang Chen , Yang Xiang , Feng Liu , Yanhua Wu , Xu Li","doi":"10.1016/j.asoc.2025.114400","DOIUrl":"10.1016/j.asoc.2025.114400","url":null,"abstract":"<div><div>To address the low detection accuracy of small pepper flower targets in complex field environments and the need for extensive annotated datasets, a refined object detection model, YOLOv5n_Pepper Flower, based on an improved YOLOv5n architecture and transfer learning, is proposed in this study. First, to enhance detection accuracy and reduce the parameter count, the conventional convolution unit within the Neck layer of the YOLOv5n model was replaced with the simplified Grouped and Shuffled Convolution (GSConv). Second, the lightweight upsampling mechanism Content Aware Feature ReAssembly (CARAFE) was integrated into the Neck layer, significantly improving accuracy with minimal additional parameters. Third, the original PANet was replaced by the Bidirectional Feature Pyramid Network (BiFPN) structure, enhancing the fusion of multi-scale features through feature weighting. Lastly, the introduction of the attention mechanism BiFormer (Vision Transformer with Bi-Level Routing Attention) into the Neck layer improved the detection capability for small pepper flower targets, consequently boosting overall model performance. Experimental results showed that the refined YOLOv5n_Pepper Flower achieved a mean average precision (mAP) of 97 %, surpassing YOLOv5n, YOLOv5s, YOLOv7tiny, YOLOv8s, and YOLOv8n by margins of 0.9 %, 0.7 %, 5.9 %, 1.3 %, and 1.7 % respectively. Additionally, the comparison between the refined YOLOv5n_Pepper Flower and YOLOv9 indicated similar detection accuracies. However, the YOLOv5n-based model exhibits superiority in terms of computation, the number of parameters, and model size, and its lightweight model was superior regarding computational efficiency, parameter count, and model size. Its lightweight characteristics are more suitable for mobile terminal deployment. Furthermore, training the refined YOLOv5n model with transfer learning increased the mAP to 97.1 %, accelerated model convergence, and improved detection performance. The findings of this study provide a solid foundation for developing methods for pepper flower posture estimation and mechanical pollination.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"188 ","pages":"Article 114400"},"PeriodicalIF":6.6,"publicationDate":"2025-12-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145749309","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

HFL-YOLOv8: a hyperbolic feature-enhanced lightweight network for object detection in ground penetrating radar images HFL-YOLOv8：用于探地雷达图像中目标检测的双曲线特征增强轻量级网络

IF 6.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Applied Soft Computing

Pub Date : 2025-12-06 DOI: 10.1016/j.asoc.2025.114403

Wentai Lei , Shixuan Yu , Tao Zhang

Deep learning-based soft computing object detection algorithms for Ground Penetrating Radar (GPR) B-Scan images face two significant challenges: Existing methods struggle to account for the unique geometric characteristics of hyperbolas. This limitation hampers the extraction of correlation information, particularly when detecting smaller objects; The high resource consumption of these algorithms increases hardware requirements, making them less efficient for practical deployment. To address these issues, this paper proposes HFL-YOLOv8, a hyperbolic feature-enhanced lightweight object detection network based on YOLOv8. The key contributions of HFL-YOLOv8 include: Three convolution layers with varying dilation rates are employed to capture feature information across different scales, improving the network’s ability to handle diverse object sizes. A dynamic upsampling operator and a channel-position attention module are incorporated to refine the detection of hyperbolic features, addressing the limitations in geometric representation. A detection head with shared parameter convolution is used to reduce computational overhead. Additionally, enhanced Sobel convolution and group normalization convolution compensate for accuracy loss, ensuring robust detection. HFL-YOLOv8 achieves an F1-score of 79.7 % and mAP50 of 81.8 %, representing significant performance improvements. Moreover, the proposed network reduces the number of parameters by 14.8 % and computational costs by 13.6 %, offering enhanced accuracy and resource efficiency for detecting small object bisectors in GPR B-Scan images.

用于探地雷达（GPR） b扫描图像的基于深度学习的软计算目标检测算法面临两个重大挑战：现有方法难以解释双曲线的独特几何特征。这种限制阻碍了相关信息的提取，特别是在检测较小的物体时；这些算法的高资源消耗增加了硬件需求，使它们在实际部署时效率较低。为了解决这些问题，本文提出了基于YOLOv8的双曲线特征增强的轻量级目标检测网络HFL-YOLOv8。HFL-YOLOv8的主要贡献包括：采用三个不同膨胀率的卷积层来捕获不同尺度的特征信息，提高了网络处理不同对象大小的能力。采用动态上采样算子和通道位置注意模块来改进双曲特征的检测，解决几何表示的局限性。采用具有共享参数卷积的检测头来减少计算开销。此外，增强的Sobel卷积和群归一化卷积补偿了精度损失，确保了鲁棒性检测。HFL-YOLOv8的f1得分为79.7%，mAP50得分为81.8%，表现出显著的性能改进。此外，该网络减少了14.8%的参数数量和13.6%的计算成本，提高了GPR b扫描图像中小目标平分线的检测精度和资源效率。

{"title":"HFL-YOLOv8: a hyperbolic feature-enhanced lightweight network for object detection in ground penetrating radar images","authors":"Wentai Lei , Shixuan Yu , Tao Zhang","doi":"10.1016/j.asoc.2025.114403","DOIUrl":"10.1016/j.asoc.2025.114403","url":null,"abstract":"<div><div>Deep learning-based soft computing object detection algorithms for Ground Penetrating Radar (GPR) B-Scan images face two significant challenges: Existing methods struggle to account for the unique geometric characteristics of hyperbolas. This limitation hampers the extraction of correlation information, particularly when detecting smaller objects; The high resource consumption of these algorithms increases hardware requirements, making them less efficient for practical deployment. To address these issues, this paper proposes HFL-YOLOv8, a hyperbolic feature-enhanced lightweight object detection network based on YOLOv8. The key contributions of HFL-YOLOv8 include: Three convolution layers with varying dilation rates are employed to capture feature information across different scales, improving the network’s ability to handle diverse object sizes. A dynamic upsampling operator and a channel-position attention module are incorporated to refine the detection of hyperbolic features, addressing the limitations in geometric representation. A detection head with shared parameter convolution is used to reduce computational overhead. Additionally, enhanced Sobel convolution and group normalization convolution compensate for accuracy loss, ensuring robust detection. HFL-YOLOv8 achieves an F1-score of 79.7 % and mAP50 of 81.8 %, representing significant performance improvements. Moreover, the proposed network reduces the number of parameters by 14.8 % and computational costs by 13.6 %, offering enhanced accuracy and resource efficiency for detecting small object bisectors in GPR B-Scan images.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"188 ","pages":"Article 114403"},"PeriodicalIF":6.6,"publicationDate":"2025-12-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145746980","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

BARL: Batch active regression learning via query-by-committee with application to slope factor of safety prediction BARL：基于委员会查询的批量主动回归学习，应用于边坡安全系数预测

IF 6.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Applied Soft Computing

Pub Date : 2025-12-06 DOI: 10.1016/j.asoc.2025.114412

Shayan Mazloom , Abdollah Shafieezadeh , Jieun Hur , Jae-Wook Jung , Jeong-Gon Ha , Daegi Hahm

Slope failure threatens critical infrastructure, requiring accurate yet efficient stability analysis for real-time monitoring and probabilistic decision-making. However, high-fidelity numerical simulations are computationally prohibitive for many-query applications, e.g., reliability analysis. While surrogate models offer an efficient alternative, they often lack accuracy and generalization across diverse geotechnical conditions. This paper introduces three key contributions to address these limitations. First, we propose Batch-Based Active Regression Learning (BARL), a novel active learning framework that extends the Query-by-Committee concept to regression tasks through a variance- and range-aware disagreement score. This new formulation quantifies predictive uncertainty and identifies the most informative samples for labeling. Second, to improve physical fidelity and generalization, our surrogate model incorporates often-overlooked parameters, namely soil saturation and porosity, and expands the input parameter space to cover a much broader range of slope geometries, initial conditions, and soil properties than those considered in prior studies. Third, we establish a new benchmark dataset derived from a high-fidelity computational model validated against a real-world landslide. This dataset is designed for regression, enabling direct prediction of continuous Factor of Safety (FoS) rather than classification. A committee of machine learning models trained with BARL demonstrated superior performance, achieving a 75 % reduction in test MSE, a 60 % reduction in test MAE, and a final test

R^{2}

of over 0.97. SHAP analysis confirmed physically consistent behaviors, establishing BARL as a robust, uncertainty-aware framework for developing accurate and efficient surrogate models for large-scale geotechnical risk assessment.

边坡破坏威胁着关键的基础设施，需要准确而有效的稳定性分析来进行实时监测和概率决策。然而，高保真度的数值模拟在计算上对许多查询应用来说是令人望而却步的，例如，可靠性分析。虽然替代模型提供了有效的替代方案，但它们在不同的岩土条件下往往缺乏准确性和泛化性。本文介绍了解决这些限制的三个关键贡献。首先，我们提出了基于批处理的主动回归学习（BARL），这是一种新的主动学习框架，它通过方差和范围感知的分歧评分将按委员会查询的概念扩展到回归任务。这个新的公式量化预测的不确定性，并确定最有信息的样本标签。其次，为了提高物理保真度和泛化，我们的代理模型包含了经常被忽视的参数，即土壤饱和度和孔隙度，并扩展了输入参数空间，以覆盖比先前研究中考虑的更广泛的斜坡几何形状、初始条件和土壤性质。第三，我们建立了一个新的基准数据集，该数据集来源于针对现实世界山体滑坡验证的高保真计算模型。该数据集是为回归设计的，可以直接预测连续的安全系数（FoS），而不是分类。使用BARL训练的机器学习模型委员会表现出优异的性能，测试MSE降低了75%，测试MAE降低了60%，最终测试R2超过0.97。SHAP分析证实了物理上一致的行为，将BARL建立为一个强大的、具有不确定性意识的框架，用于开发准确、高效的大规模岩土工程风险评估替代模型。

{"title":"BARL: Batch active regression learning via query-by-committee with application to slope factor of safety prediction","authors":"Shayan Mazloom , Abdollah Shafieezadeh , Jieun Hur , Jae-Wook Jung , Jeong-Gon Ha , Daegi Hahm","doi":"10.1016/j.asoc.2025.114412","DOIUrl":"10.1016/j.asoc.2025.114412","url":null,"abstract":"<div><div>Slope failure threatens critical infrastructure, requiring accurate yet efficient stability analysis for real-time monitoring and probabilistic decision-making. However, high-fidelity numerical simulations are computationally prohibitive for many-query applications, e.g., reliability analysis. While surrogate models offer an efficient alternative, they often lack accuracy and generalization across diverse geotechnical conditions. This paper introduces three key contributions to address these limitations. First, we propose Batch-Based Active Regression Learning (BARL), a novel active learning framework that extends the Query-by-Committee concept to regression tasks through a variance- and range-aware disagreement score. This new formulation quantifies predictive uncertainty and identifies the most informative samples for labeling. Second, to improve physical fidelity and generalization, our surrogate model incorporates often-overlooked parameters, namely soil saturation and porosity, and expands the input parameter space to cover a much broader range of slope geometries, initial conditions, and soil properties than those considered in prior studies. Third, we establish a new benchmark dataset derived from a high-fidelity computational model validated against a real-world landslide. This dataset is designed for regression, enabling direct prediction of continuous Factor of Safety (FoS) rather than classification. A committee of machine learning models trained with BARL demonstrated superior performance, achieving a 75 % reduction in test MSE, a 60 % reduction in test MAE, and a final test <span><math><msup><mi>R</mi><mn>2</mn></msup></math></span> of over 0.97. SHAP analysis confirmed physically consistent behaviors, establishing BARL as a robust, uncertainty-aware framework for developing accurate and efficient surrogate models for large-scale geotechnical risk assessment.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"188 ","pages":"Article 114412"},"PeriodicalIF":6.6,"publicationDate":"2025-12-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145749311","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

Unsupervised anomaly detection for medical image classification using masked denoising autoencoder 基于掩模去噪自编码器的医学图像分类无监督异常检测

IF 6.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Applied Soft Computing

Pub Date : 2025-12-06 DOI: 10.1016/j.asoc.2025.114388

Guangcan Qu , Haotian Pan , Chen Yang , Yezhi Lin

The application of artificial intelligence in medical image analysis is frequently hindered by the substantial cost of annotations. For diseases with low incidence rates, the imbalance between the number of cases and normal instances also presents a significant challenge to the efficacy of traditional algorithms. Unsupervised anomaly detection (UAD), which eliminates the need for labels or abnormal data during training, offers a promising solution to these issues. Consequently, we propose an advanced anomaly detection algorithm based on reconstruction techniques, termed the Masked Denoising AutoEncoder (Md-AE), for medical image classification. This approach leverages an autoencoder architecture with noise-augmentation techniques. We then introduce a masked convolution with a hybrid self-attention module, which amalgamates a channel-spatial self-attention mechanism to apply pixel-level weighting on feature maps, thereby enhancing the model’s learning capacity. Empirical results from two public datasets and one private dataset confirm that our method achieves superior classification performance. The outcomes on the private dataset underscore that UAD can be effectively employed for data that are indistinguishable by non-experts.

人工智能在医学图像分析中的应用经常受到大量注释成本的阻碍。对于低发病率的疾病，病例数与正常病例数之间的不平衡也对传统算法的有效性提出了重大挑战。无监督异常检测（UAD）在训练过程中消除了对标签或异常数据的需要，为解决这些问题提供了一个很有希望的解决方案。因此，我们提出了一种基于重建技术的高级异常检测算法，称为蒙面去噪自动编码器（Md-AE），用于医学图像分类。这种方法利用了带有噪声增强技术的自动编码器架构。然后，我们引入了一个带有混合自注意模块的掩膜卷积，该模块结合了通道-空间自注意机制，在特征映射上应用像素级加权，从而增强了模型的学习能力。两个公共数据集和一个私有数据集的实证结果证实了我们的方法取得了优异的分类性能。私有数据集的结果强调，UAD可以有效地用于非专家无法区分的数据。

{"title":"Unsupervised anomaly detection for medical image classification using masked denoising autoencoder","authors":"Guangcan Qu , Haotian Pan , Chen Yang , Yezhi Lin","doi":"10.1016/j.asoc.2025.114388","DOIUrl":"10.1016/j.asoc.2025.114388","url":null,"abstract":"<div><div>The application of artificial intelligence in medical image analysis is frequently hindered by the substantial cost of annotations. For diseases with low incidence rates, the imbalance between the number of cases and normal instances also presents a significant challenge to the efficacy of traditional algorithms. Unsupervised anomaly detection (UAD), which eliminates the need for labels or abnormal data during training, offers a promising solution to these issues. Consequently, we propose an advanced anomaly detection algorithm based on reconstruction techniques, termed the Masked Denoising AutoEncoder (Md-AE), for medical image classification. This approach leverages an autoencoder architecture with noise-augmentation techniques. We then introduce a masked convolution with a hybrid self-attention module, which amalgamates a channel-spatial self-attention mechanism to apply pixel-level weighting on feature maps, thereby enhancing the model’s learning capacity. Empirical results from two public datasets and one private dataset confirm that our method achieves superior classification performance. The outcomes on the private dataset underscore that UAD can be effectively employed for data that are indistinguishable by non-experts.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"188 ","pages":"Article 114388"},"PeriodicalIF":6.6,"publicationDate":"2025-12-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145749344","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

GCDTA: Graph-attention-assisted contrastive learning for drug-target affinity prediction GCDTA：图注意辅助对比学习用于药物靶点亲和力预测

IF 6.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Applied Soft Computing

Pub Date : 2025-12-06 DOI: 10.1016/j.asoc.2025.114376

Anqi Huang , Xiaoliang Zhou , Yupeng Wang , Ning Yu

Accurately quantifying Drug–Target Affinity (DTA) remains a fundamental yet unresolved problem in computational pharmacology. Most existing approaches process drug and protein sequences independently and subsequently concatenate their embeddings in a rudimentary fashion, thereby neglecting both the atom-level topology of molecules and richer cross-modality feature interactions. To address these limitations, we propose a novel method called Graph-Attention-Assisted Contrastive Learning for Drug-Target Affinity Prediction (GCDTA). The key idea behind GCDTA is the integration of graph-based structural representations and contrastive learning, enhancing the expressiveness and discriminative capability of learned features. Specifically, GCDTA jointly leverages both structural and sequential information. Molecular graphs of drugs are encoded using Graph Attention Networks (GAT), effectively capturing atom-level topology and local chemical environments. Protein sequences are processed using dilated convolutional layers to capture long-range dependencies and extract essential sequential features. An interactive fusion module is then employed to systematically integrate information from both modalities, enabling more effective and fine-grained cross-modal interactions. Additionally, a tailored contrastive learning objective aligns structurally analogous drug-target pairs, enhancing the discriminative power and robustness of the unified embeddings. Extensive experiments on multiple public benchmark datasets demonstrate that GCDTA consistently surpasses current state-of-the-art models, establishing new performance baselines for DTA prediction.

准确定量药物靶标亲和力（DTA）是计算药理学中一个基本但尚未解决的问题。大多数现有的方法独立处理药物和蛋白质序列，然后以一种基本的方式将它们的嵌入连接起来，从而忽略了分子的原子水平拓扑结构和更丰富的跨模态特征相互作用。为了解决这些限制，我们提出了一种新的方法，称为图-注意辅助对比学习药物-靶标亲和力预测（GCDTA）。GCDTA的核心思想是将基于图的结构表示与对比学习相结合，增强学习特征的表达能力和判别能力。具体来说，GCDTA同时利用了结构信息和顺序信息。使用图形注意网络（GAT）对药物分子图进行编码，有效捕获原子级拓扑结构和局部化学环境。使用扩展卷积层处理蛋白质序列以捕获远程依赖关系并提取基本序列特征。然后使用交互式融合模块系统地集成来自两种模式的信息，从而实现更有效和细粒度的跨模式交互。此外，定制的对比学习目标将结构类似的药物-目标对对齐，增强了统一嵌入的判别能力和鲁棒性。在多个公共基准数据集上进行的大量实验表明，GCDTA始终优于当前最先进的模型，为DTA预测建立了新的性能基准。

{"title":"GCDTA: Graph-attention-assisted contrastive learning for drug-target affinity prediction","authors":"Anqi Huang , Xiaoliang Zhou , Yupeng Wang , Ning Yu","doi":"10.1016/j.asoc.2025.114376","DOIUrl":"10.1016/j.asoc.2025.114376","url":null,"abstract":"<div><div>Accurately quantifying Drug–Target Affinity (DTA) remains a fundamental yet unresolved problem in computational pharmacology. Most existing approaches process drug and protein sequences independently and subsequently concatenate their embeddings in a rudimentary fashion, thereby neglecting both the atom-level topology of molecules and richer cross-modality feature interactions. To address these limitations, we propose a novel method called <strong>G</strong>raph-Attention-Assisted <strong>C</strong>ontrastive Learning for <strong>D</strong>rug-<strong>T</strong>arget <strong>A</strong>ffinity Prediction (GCDTA). The key idea behind GCDTA is the integration of graph-based structural representations and contrastive learning, enhancing the expressiveness and discriminative capability of learned features. Specifically, GCDTA jointly leverages both structural and sequential information. Molecular graphs of drugs are encoded using Graph Attention Networks (GAT), effectively capturing atom-level topology and local chemical environments. Protein sequences are processed using dilated convolutional layers to capture long-range dependencies and extract essential sequential features. An interactive fusion module is then employed to systematically integrate information from both modalities, enabling more effective and fine-grained cross-modal interactions. Additionally, a tailored contrastive learning objective aligns structurally analogous drug-target pairs, enhancing the discriminative power and robustness of the unified embeddings. Extensive experiments on multiple public benchmark datasets demonstrate that GCDTA consistently surpasses current state-of-the-art models, establishing new performance baselines for DTA prediction.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"188 ","pages":"Article 114376"},"PeriodicalIF":6.6,"publicationDate":"2025-12-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145746979","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

Unraveling incomplete multiview data: An information-theoretic framework 揭示不完整的多视图数据：一个信息论框架

IF 6.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Applied Soft Computing

Pub Date : 2025-12-06 DOI: 10.1016/j.asoc.2025.114405

Qi Zhang, Mingfei Lu, Jingmin Xin, Badong Chen

Multiview learning can exploit complementary information from multiple sources, but its effectiveness is often limited by missing views. A central challenge is to distinguish shared information across views from view-specific variations, since treating them as a single representation introduces noise and weakens generalization. To tackle this problem, we propose an information-theoretic framework that explicitly separates consistent and view-specific components. The shared representation captures essential cross-view dependencies, while unique information is retained for each view. An information bottleneck strategy further compresses both components to keep the most relevant features and reduce redundancy. For estimation, we employ a matrix-based Rényi’s

α

-order entropy functional, which allows direct optimization without variational approximations. Extensive experiments on eight real-world datasets demonstrate that our method consistently outperforms state-of-the-art baselines, with the advantage becoming more pronounced as the missing rate increases. Moreover, ablation studies confirm that each component—shared information, view-specific information, and the information bottleneck—is indispensable, and removing any of them results in a performance drop. In conclusion, explicitly modeling both shared and view-specific information within an information-theoretic framework provides a principled, robust, and generalizable solution for incomplete multiview learning. Code is available at https://github.com/archy666/UIMD.

多视图学习可以利用来自多个来源的互补信息，但其有效性往往受到缺失视图的限制。一个核心挑战是区分视图之间的共享信息和特定于视图的变化，因为将它们视为单一表示会引入噪声并削弱泛化。为了解决这个问题，我们提出了一个信息理论框架，它显式地分离一致性和特定于视图的组件。共享表示捕获基本的跨视图依赖关系，同时保留每个视图的唯一信息。信息瓶颈策略进一步压缩这两个组件，以保留最相关的特性并减少冗余。对于估计，我们采用基于矩阵的r -阶熵函数，它允许直接优化而不需要变分近似。在八个真实数据集上进行的大量实验表明，我们的方法始终优于最先进的基线，随着缺失率的增加，优势变得更加明显。此外，消融研究证实，每个组件（共享信息、特定于视图的信息和信息瓶颈）都是必不可少的，删除它们中的任何一个都会导致性能下降。总之，在信息论框架内显式地对共享信息和特定于视图的信息进行建模，为不完全多视图学习提供了一个原则性的、健壮的和可推广的解决方案。代码可从https://github.com/archy666/UIMD获得。

{"title":"Unraveling incomplete multiview data: An information-theoretic framework","authors":"Qi Zhang, Mingfei Lu, Jingmin Xin, Badong Chen","doi":"10.1016/j.asoc.2025.114405","DOIUrl":"10.1016/j.asoc.2025.114405","url":null,"abstract":"<div><div>Multiview learning can exploit complementary information from multiple sources, but its effectiveness is often limited by missing views. A central challenge is to distinguish shared information across views from view-specific variations, since treating them as a single representation introduces noise and weakens generalization. To tackle this problem, we propose an information-theoretic framework that explicitly separates consistent and view-specific components. The shared representation captures essential cross-view dependencies, while unique information is retained for each view. An information bottleneck strategy further compresses both components to keep the most relevant features and reduce redundancy. For estimation, we employ a matrix-based Rényi’s <span><math><mi>α</mi></math></span>-order entropy functional, which allows direct optimization without variational approximations. Extensive experiments on eight real-world datasets demonstrate that our method consistently outperforms state-of-the-art baselines, with the advantage becoming more pronounced as the missing rate increases. Moreover, ablation studies confirm that each component—shared information, view-specific information, and the information bottleneck—is indispensable, and removing any of them results in a performance drop. In conclusion, explicitly modeling both shared and view-specific information within an information-theoretic framework provides a principled, robust, and generalizable solution for incomplete multiview learning. <em>Code is available at</em> <span><span>https://github.com/archy666/UIMD</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"188 ","pages":"Article 114405"},"PeriodicalIF":6.6,"publicationDate":"2025-12-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145749307","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

Long-short term time-delay reservoir for nonlinear time series forecasting task 用于非线性时间序列预测任务的长-短时滞库

IF 6.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Applied Soft Computing

Pub Date : 2025-12-06 DOI: 10.1016/j.asoc.2025.114418

Dongchen Jiang , Yi Zeng , Meiming You , Guoqiang Wang

Reservoir computing (RC) such as Time-Delay Reservoir (TDR) has gained attention due to its good hardware scalability structure and low training complexity. However, most existing TDR variants follow shallow architectures, limiting their short-term memory (STM) and multi-scale modelling capabilities. To address these limitations, this paper proposes a novel wide-layered TDR framework and two multi-scale deep TDR models: Long-delay Time-Delay Reservoir (LD-TDR) and Long-short Term Time-Delay Reservoir (LS-TDR). LD-TDR integrates two independently operated DTDR modules, with skip connections and a cross-coupling mechanism to enhance long-term memory retention. LS-TDR extends LD-TDR by introducing a short-term reservoir, improving the model’s ability to capture local and short-term dynamics. Experiments are conducted on both chaotic and real-world datasets, including Mackey-Glass, Lorenz, Sunspot, ETT and network traffic datasets, under open-loop and closed-loop prediction settings. Results show that the proposed models consistently outperform the TDR variants (TDR, DDR, DTDR and DATDR) and Echo State Networks (Deep-ESN, ADRC, LS-ESN and LS-CrossESN) in both single-step and multi-step forecasting tasks. The LD-TDR and LS-TDR exhibit significantly enhanced STM and improved multi-scale temporal representation, demonstrating predictive accuracy across diverse datasets.

时延库（TDR）等库计算以其良好的硬件可扩展性结构和较低的训练复杂度而备受关注。然而，大多数现有的TDR变体遵循浅层架构，限制了它们的短期记忆（STM）和多尺度建模能力。为了解决这些问题，本文提出了一种新的宽层TDR框架和两个多尺度深度TDR模型：长延迟时滞水库（LD-TDR）和长短期时滞水库（LS-TDR）。LD-TDR集成了两个独立运行的DTDR模块，具有跳过连接和交叉耦合机制，以增强长期记忆保留。LS-TDR通过引入短期储层来扩展LD-TDR，提高了模型捕捉局部和短期动态的能力。在开环和闭环预测设置下，在混沌和现实数据集上进行了实验，包括Mackey-Glass、Lorenz、Sunspot、ETT和网络流量数据集。结果表明，该模型在单步和多步预测任务中均优于TDR变体（TDR、DDR、DTDR和DATDR）和回声状态网络（Deep-ESN、ADRC、LS-ESN和LS-CrossESN）。LD-TDR和LS-TDR表现出显著增强的STM和改进的多尺度时间表征，在不同数据集上显示出预测准确性。

{"title":"Long-short term time-delay reservoir for nonlinear time series forecasting task","authors":"Dongchen Jiang , Yi Zeng , Meiming You , Guoqiang Wang","doi":"10.1016/j.asoc.2025.114418","DOIUrl":"10.1016/j.asoc.2025.114418","url":null,"abstract":"<div><div>Reservoir computing (RC) such as Time-Delay Reservoir (TDR) has gained attention due to its good hardware scalability structure and low training complexity. However, most existing TDR variants follow shallow architectures, limiting their short-term memory (STM) and multi-scale modelling capabilities. To address these limitations, this paper proposes a novel wide-layered TDR framework and two multi-scale deep TDR models: Long-delay Time-Delay Reservoir (LD-TDR) and Long-short Term Time-Delay Reservoir (LS-TDR). LD-TDR integrates two independently operated DTDR modules, with skip connections and a cross-coupling mechanism to enhance long-term memory retention. LS-TDR extends LD-TDR by introducing a short-term reservoir, improving the model’s ability to capture local and short-term dynamics. Experiments are conducted on both chaotic and real-world datasets, including Mackey-Glass, Lorenz, Sunspot, ETT and network traffic datasets, under open-loop and closed-loop prediction settings. Results show that the proposed models consistently outperform the TDR variants (TDR, DDR, DTDR and DATDR) and Echo State Networks (Deep-ESN, ADRC, LS-ESN and LS-CrossESN) in both single-step and multi-step forecasting tasks. The LD-TDR and LS-TDR exhibit significantly enhanced STM and improved multi-scale temporal representation, demonstrating predictive accuracy across diverse datasets.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"188 ","pages":"Article 114418"},"PeriodicalIF":6.6,"publicationDate":"2025-12-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145746978","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 probability-based linguistic decision-making approach to a reliable Kano model for classifying quality attributes 一种基于概率的语言决策方法，用于质量属性分类的可靠Kano模型

IF 6.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Applied Soft Computing

Pub Date : 2025-12-06 DOI: 10.1016/j.asoc.2025.114381

Hong-Bin Yan , Lianzhuang Qu , Van-Nam Huynh

Due to its ability to reflect the asymmetric and nonlinear relationship between quality and customer satisfaction, the Kano model has been widely practiced in marketing and product/service design for classifying quality attributes. Despite the methodological revisions and extensions of the Kano model in the literature, its effective utilization is still critically challenged by the following reliability issues: uncertainties underlying both evaluation rules and the Kano survey as well as lack of a suitable reliability measure for the Kano survey. To address these three issues simultaneously, this paper seeks to propose a probability-based linguistic decision-making approach to a reliable Kano model (PKM). To do so, a Bayesian ensemble model is first proposed to infer uncertain evaluation rules, the reliabilities of which are well validated and justified. Second, our PKM couples the inferred uncertain rules with Kano survey data in terms of precise assessments, fuzzy assessments, and hesitant fuzzy linguistic term sets, into the aggregation and choice functions. Third, a probability-based reliability measure is proposed for the special characteristics as well as precise and uncertain expressions of the Kano survey. Two comparative application studies reveal that our PKM excels in handling uncertainties of evaluation rules, providing a richer way to reliable expressions in the Kano survey, measuring the reliability indices of the Kano survey, as well as avoiding the information loss in aggregating Kano survey data. As such, this paper provides researchers and managers a “soft” reliable Kano model for classifying quality attributes as well as sheds new light on applications of fuzzy linguistic approaches.

由于能够反映质量与顾客满意度之间的不对称和非线性关系，Kano模型在市场营销和产品/服务设计中被广泛应用于质量属性分类。尽管文献中对Kano模型进行了方法学上的修订和扩展，但其有效利用仍然受到以下可靠性问题的严峻挑战：评估规则和Kano调查的不确定性以及缺乏适合Kano调查的可靠性度量。为了同时解决这三个问题，本文试图提出一种基于概率的语言决策方法来实现可靠的Kano模型（PKM）。为此，首先提出了一个贝叶斯集成模型来推断不确定的评估规则，其可靠性得到了很好的验证和证明。其次，我们的PKM将推断的不确定规则与Kano调查数据在精确评估、模糊评估和犹豫模糊语言术语集方面耦合到聚合和选择函数中。第三，针对卡诺测量的特殊性以及其精确和不确定的表达，提出了一种基于概率的可靠性度量方法。两项对比应用研究表明，我们的PKM在处理评价规则的不确定性方面表现出色，为卡诺调查的可靠性表达提供了更丰富的途径，测量了卡诺调查的可靠性指标，并避免了卡诺调查数据汇总时的信息丢失。因此，本文为研究人员和管理人员提供了一个“软”可靠的Kano模型来分类质量属性，并为模糊语言方法的应用提供了新的思路。

{"title":"A probability-based linguistic decision-making approach to a reliable Kano model for classifying quality attributes","authors":"Hong-Bin Yan , Lianzhuang Qu , Van-Nam Huynh","doi":"10.1016/j.asoc.2025.114381","DOIUrl":"10.1016/j.asoc.2025.114381","url":null,"abstract":"<div><div>Due to its ability to reflect the asymmetric and nonlinear relationship between quality and customer satisfaction, the Kano model has been widely practiced in marketing and product/service design for classifying quality attributes. Despite the methodological revisions and extensions of the Kano model in the literature, its effective utilization is still critically challenged by the following reliability issues: uncertainties underlying both evaluation rules and the Kano survey as well as lack of a suitable reliability measure for the Kano survey. To address these three issues simultaneously, this paper seeks to propose a probability-based linguistic decision-making approach to a reliable Kano model (PKM). To do so, a Bayesian ensemble model is first proposed to infer uncertain evaluation rules, the reliabilities of which are well validated and justified. Second, our PKM couples the inferred uncertain rules with Kano survey data in terms of precise assessments, fuzzy assessments, and hesitant fuzzy linguistic term sets, into the aggregation and choice functions. Third, a probability-based reliability measure is proposed for the special characteristics as well as precise and uncertain expressions of the Kano survey. Two comparative application studies reveal that our PKM excels in handling uncertainties of evaluation rules, providing a richer way to reliable expressions in the Kano survey, measuring the reliability indices of the Kano survey, as well as avoiding the information loss in aggregating Kano survey data. As such, this paper provides researchers and managers a “soft” reliable Kano model for classifying quality attributes as well as sheds new light on applications of fuzzy linguistic approaches.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"188 ","pages":"Article 114381"},"PeriodicalIF":6.6,"publicationDate":"2025-12-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145749348","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

Enhancing evolutionary algorithms with solution prediction for influence maximization in social networks 社会网络中影响最大化的改进进化算法与解预测

IF 6.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Applied Soft Computing

Pub Date : 2025-12-06 DOI: 10.1016/j.asoc.2025.114399

Kaicong Ma , Haipeng Yang , Hangxing Ma , Xinxiang Xu , Qiang He , Lei Zhang

Influence Maximization (IM) aims to select a subset of users from a social network to maximize the influence spread. Due to the NP-hardness of the IM problem, most existing IM methods either suffer from high computational complexity or compromise solution accuracy. To address this challenge, we propose a Classification-assisted Evolutionary Algorithm (CEA), which leverages the information from solution prediction to boost the Evolutionary Algorithm (EA) to efficiently and effectively address the IM problem. In the first place, we train a classification model using a group of small IM problem instances with given optimal solutions. Subsequently, the offline-trained classification model is utilized to predict whether a node belongs to the optimal solution or not, and further predicts the probability that it is part of the optimal solution. In the second place, the information from solution prediction is integrated into the key components of EA to facilitate the search for high-quality solutions, including search space partition, initialization, and evolutionary operators. Extensive experiments conducted on 12 real-world social networks demonstrate that our CEA algorithm achieves an average of 99 % of the influence spread achieved by cost-effective lazy forward (CELF), while reducing running time by at least two orders of magnitude in most cases. Consequently, CEA strikes a better balance between effectiveness and efficiency.

影响力最大化（IM）旨在从社交网络中选择一部分用户，以最大限度地扩大影响力传播。由于IM问题的np -硬度，现有的IM方法要么计算量大，要么求解精度低。为了解决这一挑战，我们提出了一种分类辅助进化算法（CEA），该算法利用解决方案预测的信息来增强进化算法（EA），以高效地解决IM问题。首先，我们使用一组具有给定最优解的小型IM问题实例来训练分类模型。随后，利用离线训练的分类模型预测节点是否属于最优解，并进一步预测其属于最优解的概率。其次，将来自解决方案预测的信息集成到EA的关键组件中，以促进对高质量解决方案的搜索，包括搜索空间划分、初始化和进化操作符。在12个真实社会网络上进行的大量实验表明，我们的CEA算法平均达到了成本效益延迟前向（CELF）所达到的影响传播的99%，同时在大多数情况下将运行时间减少了至少两个数量级。因此，CEA在有效性和效率之间取得了更好的平衡。

{"title":"Enhancing evolutionary algorithms with solution prediction for influence maximization in social networks","authors":"Kaicong Ma , Haipeng Yang , Hangxing Ma , Xinxiang Xu , Qiang He , Lei Zhang","doi":"10.1016/j.asoc.2025.114399","DOIUrl":"10.1016/j.asoc.2025.114399","url":null,"abstract":"<div><div>Influence Maximization (IM) aims to select a subset of users from a social network to maximize the influence spread. Due to the NP-hardness of the IM problem, most existing IM methods either suffer from high computational complexity or compromise solution accuracy. To address this challenge, we propose a Classification-assisted Evolutionary Algorithm (CEA), which leverages the information from solution prediction to boost the Evolutionary Algorithm (EA) to efficiently and effectively address the IM problem. In the first place, we train a classification model using a group of small IM problem instances with given optimal solutions. Subsequently, the offline-trained classification model is utilized to predict whether a node belongs to the optimal solution or not, and further predicts the probability that it is part of the optimal solution. In the second place, the information from solution prediction is integrated into the key components of EA to facilitate the search for high-quality solutions, including search space partition, initialization, and evolutionary operators. Extensive experiments conducted on 12 real-world social networks demonstrate that our CEA algorithm achieves an average of 99 % of the influence spread achieved by cost-effective lazy forward (CELF), while reducing running time by at least two orders of magnitude in most cases. Consequently, CEA strikes a better balance between effectiveness and efficiency.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"188 ","pages":"Article 114399"},"PeriodicalIF":6.6,"publicationDate":"2025-12-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145749345","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