Intelligent Systems with Applications最新文献_第5页

Fusing explainable deep learning ensembles and LLM recommendations for real-time plant leaf disease diagnosis 融合可解释的深度学习集合和LLM建议用于实时植物叶片疾病诊断

IF 4.3

Intelligent Systems with Applications

Pub Date : 2025-10-24 DOI: 10.1016/j.iswa.2025.200596

Dip Kumar Saha , Mohammad Rasel Ahmed , Tushar Deb Nath , Rounakul Islam Boby , Md. Jakir Hossen , M.F. Mridha

Timely and accurate identification of plant leaf diseases plays a vital role in ensuring sustainable agriculture and universal food security. Accurate identification of plant leaf diseases ensures healthier plant cultivation, which is pivotal for sustainable agriculture operations. In this study, we present a plant leaf disease recognition mechanism that utilizes a stacking ensemble structure combined with a Large Language Model (LLM) and Explainable AI (XAI) mechanism to improve identification accuracy and comprehensibility. To capture high textural structure, we utilized the Gray Level Co-occurrence Matrix (GLCM), whereas the MobileNetV3 architecture was utilized to maintain low computational cost in feature extraction. GoogleNet was integrated to improve multi-scale feature extraction by employing inception blocks, which effectively obtain fine-grained details and universal spatial patterns. Our ensemble framework integrates improved versions of MobileNetV3, GoogleNet, and ConvNeXtSmall with CatBoost employed as a nonlinear meta-learner allowing the framework to effectively capture complex connections among the base models within the ensemble framework. Moreover, we utilized additional CNN models, including AlexNet and EfficientNetV2B0, to compare the result of our proposed stacking ensemble model and to evaluate its generalization ability over various architectural designs. In addition, we developed a real-time system integrating an LLM with the proposed ensemble model, ensuring automatic plant leaf disease recognition and delivering corresponding curing recommendations. Our findings contribute to plant-based agriculture by enabling early diagnosis of leaf diseases and providing real-time recommendations through DL and LLM technology.

及时准确地识别植物叶片病害对确保可持续农业和普遍粮食安全具有至关重要的作用。准确识别植物叶片病害可确保更健康的植物种植，这对可持续农业经营至关重要。在本研究中，我们提出了一种植物叶片病害识别机制，该机制利用堆叠集成结构结合大语言模型（Large Language Model， LLM）和可解释人工智能（explable AI， XAI）机制来提高识别精度和可理解性。为了捕获高纹理结构，我们使用了灰度共生矩阵（GLCM），而在特征提取中，我们使用了MobileNetV3架构来保持较低的计算成本。集成GoogleNet，利用初始块改进多尺度特征提取，有效获取细粒度细节和通用空间模式。我们的集成框架集成了MobileNetV3、GoogleNet和ConvNeXtSmall的改进版本，并使用CatBoost作为非线性元学习器，允许框架有效地捕获集成框架内基本模型之间的复杂连接。此外，我们使用了额外的CNN模型，包括AlexNet和EfficientNetV2B0，来比较我们提出的堆叠集成模型的结果，并评估其在各种建筑设计上的泛化能力。此外，我们还开发了一个实时系统，将LLM与所提出的集成模型集成在一起，确保植物叶片病害的自动识别并提供相应的养护建议。我们的研究结果有助于植物农业，通过DL和LLM技术实现叶片疾病的早期诊断并提供实时建议。

{"title":"Fusing explainable deep learning ensembles and LLM recommendations for real-time plant leaf disease diagnosis","authors":"Dip Kumar Saha , Mohammad Rasel Ahmed , Tushar Deb Nath , Rounakul Islam Boby , Md. Jakir Hossen , M.F. Mridha","doi":"10.1016/j.iswa.2025.200596","DOIUrl":"10.1016/j.iswa.2025.200596","url":null,"abstract":"<div><div>Timely and accurate identification of plant leaf diseases plays a vital role in ensuring sustainable agriculture and universal food security. Accurate identification of plant leaf diseases ensures healthier plant cultivation, which is pivotal for sustainable agriculture operations. In this study, we present a plant leaf disease recognition mechanism that utilizes a stacking ensemble structure combined with a Large Language Model (LLM) and Explainable AI (XAI) mechanism to improve identification accuracy and comprehensibility. To capture high textural structure, we utilized the Gray Level Co-occurrence Matrix (GLCM), whereas the MobileNetV3 architecture was utilized to maintain low computational cost in feature extraction. GoogleNet was integrated to improve multi-scale feature extraction by employing inception blocks, which effectively obtain fine-grained details and universal spatial patterns. Our ensemble framework integrates improved versions of MobileNetV3, GoogleNet, and ConvNeXtSmall with CatBoost employed as a nonlinear meta-learner allowing the framework to effectively capture complex connections among the base models within the ensemble framework. Moreover, we utilized additional CNN models, including AlexNet and EfficientNetV2B0, to compare the result of our proposed stacking ensemble model and to evaluate its generalization ability over various architectural designs. In addition, we developed a real-time system integrating an LLM with the proposed ensemble model, ensuring automatic plant leaf disease recognition and delivering corresponding curing recommendations. Our findings contribute to plant-based agriculture by enabling early diagnosis of leaf diseases and providing real-time recommendations through DL and LLM technology.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"28 ","pages":"Article 200596"},"PeriodicalIF":4.3,"publicationDate":"2025-10-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145416657","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Authenticating devices based on audio feature selection with scene-specific tuning and landmark augmentation 基于音频特征选择与场景特定的调谐和地标增强验证设备

IF 4.3

Intelligent Systems with Applications

Pub Date : 2025-10-15 DOI: 10.1016/j.iswa.2025.200593

Andi Bahtiar Semma , Kusrini , Arief Setyanto , Bruno da Silva , An Braeken

Authentication methods have evolved significantly, transitioning from traditional passwords to biometric and multi-factor techniques, with audio-based systems now emerging as a promising frontier. These systems leverage ambient sounds or device-generated noise for continuous authentication but encounter challenges such as environmental noise, spoofing risks, and standardized audio feature selection. This study tackles these issues by focusing on robust handling of environmental variations and interference and optimizing audio feature selection for effective environmental audio authentication. A key innovation introduced is the concept of “audio landmarks,” randomly generated signals embedded into audio samples. These landmarks enhance device authentication by enriching feature representation and reducing sensitivity to noise, leading to significant improvements in precision, recall, and F1 scores across various scenarios. In some cases, features achieved a perfect F1 score 1.00 under ideal conditions. Among audio features analyzed, the Constant-Q Transform (CQT) excels, particularly in music or speech scenes. However, combining multiple features often introduces redundancy due to overlapping information and varying optimal thresholds, which may not constantly improve performance. Additionally, spectral centroids and spectral contrast, which are computationally lightweight at 9 ms and 10 ms, respectively, deliver excellent performance, making them ideal for real-time or resource-constrained applications, as tested on the Raspberry Pi 4. These findings provide practical guidelines for audio-based device authentication by leveraging cryptographic hashing for deterministic landmark generation and the balanced fusion of landmark and acoustic features. This enables robust authentication even in challenging scenarios where environmental sounds are insufficient.

身份验证方法已经发生了重大变化，从传统的密码过渡到生物识别和多因素技术，基于音频的系统现在正在成为一个有前途的前沿。这些系统利用环境声音或设备产生的噪声进行持续身份验证，但会遇到环境噪声、欺骗风险和标准化音频特征选择等挑战。本研究通过关注环境变化和干扰的鲁棒处理以及优化音频特征选择以实现有效的环境音频认证来解决这些问题。引入的一个关键创新是“音频地标”概念，即嵌入音频样本中随机生成的信号。这些标志通过丰富特征表示和降低对噪声的敏感性来增强设备认证，从而显著提高了各种场景下的精度、召回率和F1分数。在某些情况下，功能在理想条件下获得了完美的F1分数1.00。在分析的音频特征中，恒定q变换（CQT）表现出色，特别是在音乐或语音场景中。然而，组合多个特征往往会由于信息重叠和最优阈值变化而引入冗余，这可能不会持续提高性能。此外，光谱质心和光谱对比度，分别在9毫秒和10毫秒的计算轻量级，提供了出色的性能，使它们成为实时或资源受限应用的理想选择，正如在Raspberry Pi 4上测试的那样。这些发现为基于音频的设备认证提供了实用指南，通过利用加密散列来确定地标生成以及地标和声学特征的平衡融合。即使在环境声音不足的挑战性场景中，这也可以实现健壮的身份验证。

{"title":"Authenticating devices based on audio feature selection with scene-specific tuning and landmark augmentation","authors":"Andi Bahtiar Semma , Kusrini , Arief Setyanto , Bruno da Silva , An Braeken","doi":"10.1016/j.iswa.2025.200593","DOIUrl":"10.1016/j.iswa.2025.200593","url":null,"abstract":"<div><div>Authentication methods have evolved significantly, transitioning from traditional passwords to biometric and multi-factor techniques, with audio-based systems now emerging as a promising frontier. These systems leverage ambient sounds or device-generated noise for continuous authentication but encounter challenges such as environmental noise, spoofing risks, and standardized audio feature selection. This study tackles these issues by focusing on robust handling of environmental variations and interference and optimizing audio feature selection for effective environmental audio authentication. A key innovation introduced is the concept of “audio landmarks,” randomly generated signals embedded into audio samples. These landmarks enhance device authentication by enriching feature representation and reducing sensitivity to noise, leading to significant improvements in precision, recall, and F1 scores across various scenarios. In some cases, features achieved a perfect F1 score 1.00 under ideal conditions. Among audio features analyzed, the Constant-Q Transform (CQT) excels, particularly in music or speech scenes. However, combining multiple features often introduces redundancy due to overlapping information and varying optimal thresholds, which may not constantly improve performance. Additionally, spectral centroids and spectral contrast, which are computationally lightweight at 9 ms and 10 ms, respectively, deliver excellent performance, making them ideal for real-time or resource-constrained applications, as tested on the Raspberry Pi 4. These findings provide practical guidelines for audio-based device authentication by leveraging cryptographic hashing for deterministic landmark generation and the balanced fusion of landmark and acoustic features. This enables robust authentication even in challenging scenarios where environmental sounds are insufficient.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"28 ","pages":"Article 200593"},"PeriodicalIF":4.3,"publicationDate":"2025-10-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145321013","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

AGU: Adaptive gradient unlearning for efficient machine unlearning AGU：用于高效机器学习的自适应梯度学习

IF 4.3

Intelligent Systems with Applications

Pub Date : 2025-10-07 DOI: 10.1016/j.iswa.2025.200592

Naglaa E. Ghannam , Esraa A. Mahareek

The right to be forgotten is an essential requirement for machine learning systems under privacy regulations such as the GDPR and CCPA. We introduce Adaptive Gradient Unlearning (AGU), a novel influence-based algorithm designed to efficiently remove the contribution of specified training data while preserving overall model utility. Unlike retraining-based methods, AGU calculates parameter-level gradient sensitivity scores over the forget set to identify which weights are most influenced by the data targeted for deletion. These scores are then normalized and used to adaptively scale gradient updates, selectively erasing data influence without disrupting unrelated knowledge. Convergence is managed via dual stopping criteria based on changes in model parameters and empirical privacy leakage, which is measured by prediction divergence before and after unlearning. AGU achieves strong empirical results on six benchmark datasets including MNIST, CIFAR-10, CIFAR-100, IMDB, UCI Adult, and Tiny-ImageNet-200. In comparison with state-of-the-art methods such as SISA, SCRUB, AmnesiacML, SALUN, Boundary Unlearning, and retraining (ORTR) as a benchmark, AGU yields the best accuracy retention, unlearning times, memory overhead, and privacy leak. For example: AGU achieves an average of 98.3 % on MNIST while unlearning four times faster and using a third of the memory cost in comparison with ORTR. These results make a case for AGU as a practical, scalable data deletion approach with privacy guarantees in an era of deep learning and further extendable into federated and decentralized systems.

根据GDPR和CCPA等隐私法规，被遗忘权是机器学习系统的基本要求。我们引入了自适应梯度学习（AGU），这是一种新的基于影响的算法，旨在有效地去除指定训练数据的贡献，同时保持整体模型的实用性。与基于再训练的方法不同，AGU在遗忘集上计算参数级梯度灵敏度分数，以确定哪些权重受删除目标数据的影响最大。然后将这些分数归一化并用于自适应缩放梯度更新，在不破坏不相关知识的情况下选择性地消除数据影响。收敛通过基于模型参数变化和经验隐私泄漏的双停止准则进行管理，并通过学习前后的预测散度来衡量。AGU在MNIST、CIFAR-10、CIFAR-100、IMDB、UCI Adult、Tiny-ImageNet-200等6个基准数据集上取得了较强的实证结果。与最先进的方法（如SISA、SCRUB、AmnesiacML、SALUN、边界学习和再训练（ORTR））作为基准相比，AGU产生了最好的准确性保持、学习时间、内存开销和隐私泄漏。例如：AGU在MNIST上平均达到98.3%，而遗忘速度是ORTR的四倍，使用的内存成本是ORTR的三分之一。这些结果证明，在深度学习时代，AGU是一种实用的、可扩展的数据删除方法，具有隐私保障，并可进一步扩展到联邦和分散的系统中。

{"title":"AGU: Adaptive gradient unlearning for efficient machine unlearning","authors":"Naglaa E. Ghannam , Esraa A. Mahareek","doi":"10.1016/j.iswa.2025.200592","DOIUrl":"10.1016/j.iswa.2025.200592","url":null,"abstract":"<div><div>The right to be forgotten is an essential requirement for machine learning systems under privacy regulations such as the GDPR and CCPA. We introduce Adaptive Gradient Unlearning (AGU), a novel influence-based algorithm designed to efficiently remove the contribution of specified training data while preserving overall model utility. Unlike retraining-based methods, AGU calculates parameter-level gradient sensitivity scores over the forget set to identify which weights are most influenced by the data targeted for deletion. These scores are then normalized and used to adaptively scale gradient updates, selectively erasing data influence without disrupting unrelated knowledge. Convergence is managed via dual stopping criteria based on changes in model parameters and empirical privacy leakage, which is measured by prediction divergence before and after unlearning. AGU achieves strong empirical results on six benchmark datasets including MNIST, CIFAR-10, CIFAR-100, IMDB, UCI Adult, and Tiny-ImageNet-200. In comparison with state-of-the-art methods such as SISA, SCRUB, AmnesiacML, SALUN, Boundary Unlearning, and retraining (ORTR) as a benchmark, AGU yields the best accuracy retention, unlearning times, memory overhead, and privacy leak. For example: AGU achieves an average of 98.3 % on MNIST while unlearning four times faster and using a third of the memory cost in comparison with ORTR. These results make a case for AGU as a practical, scalable data deletion approach with privacy guarantees in an era of deep learning and further extendable into federated and decentralized systems.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"28 ","pages":"Article 200592"},"PeriodicalIF":4.3,"publicationDate":"2025-10-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145321014","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

AI safety practices and public perception: Historical analysis, survey insights, and a weighted scoring framework 人工智能安全实践和公众认知：历史分析、调查见解和加权评分框架

IF 4.3

Intelligent Systems with Applications

Pub Date : 2025-10-01 DOI: 10.1016/j.iswa.2025.200583

Maikel Leon

Artificial Intelligence (AI) safety has evolved in tandem with advances in technology and shifts in societal attitudes. This article presents a historical and empirical analysis of AI safety concerns from the mid-twentieth century to the present, integrating archival records, media narratives, survey data, landmark research, and regulatory developments. Early anxieties (rooted in Cold War geopolitics and science fiction) focused on physical robots and autonomous weapons. In contrast, contemporary debates focus on algorithmic bias, misinformation, job displacement, and existential risks posed by advanced systems, such as Large Language Models (LLMs). This article examines the impact of key scholarly contributions, significant events, and regulatory milestones on public perception and governance approaches. Building on this context, this study proposes an improved LLM safety scoring system that prioritizes existential risk mitigation, transparency, and governance accountability. Applying the proposed framework to leading AI developers reveals significant variation in safety commitments. The results underscore how weighting choices affect rankings. Comparative analysis with existing indices highlights the importance of nuanced, multidimensional evaluation methods. The paper concludes by identifying pressing governance challenges, including the need for global cooperation, robust interpretability, and ongoing monitoring of harm in high-stakes domains. These findings demonstrate that AI safety is not static but somewhat shaped by historical context, technical capabilities, and societal values—requiring the continuous adaptation of both policy and evaluation frameworks to align AI systems with human interests.

随着技术的进步和社会态度的转变，人工智能（AI）安全性也在不断发展。本文对20世纪中期至今的人工智能安全问题进行了历史和实证分析，整合了档案记录、媒体叙述、调查数据、里程碑式研究和监管发展。早期的焦虑（源于冷战地缘政治和科幻小说）集中在实体机器人和自主武器上。相比之下，当代的争论集中在算法偏见、错误信息、工作取代以及大型语言模型（llm）等先进系统带来的存在风险上。本文考察了关键学术贡献、重大事件和监管里程碑对公众认知和治理方法的影响。在此背景下，本研究提出了一个改进的法学硕士安全评分系统，优先考虑存在风险缓解、透明度和治理问责制。将提出的框架应用于领先的人工智能开发人员，会发现他们在安全承诺方面存在显著差异。结果强调了权重选择是如何影响排名的。与现有指标的对比分析凸显了细致入微、多维度评价方法的重要性。本文最后指出了紧迫的治理挑战，包括对全球合作的需求、强有力的可解释性以及对高风险领域危害的持续监测。这些发现表明，人工智能的安全性不是静态的，而是在一定程度上受到历史背景、技术能力和社会价值观的影响，需要不断调整政策和评估框架，使人工智能系统与人类利益保持一致。

{"title":"AI safety practices and public perception: Historical analysis, survey insights, and a weighted scoring framework","authors":"Maikel Leon","doi":"10.1016/j.iswa.2025.200583","DOIUrl":"10.1016/j.iswa.2025.200583","url":null,"abstract":"<div><div>Artificial Intelligence (AI) safety has evolved in tandem with advances in technology and shifts in societal attitudes. This article presents a historical and empirical analysis of AI safety concerns from the mid-twentieth century to the present, integrating archival records, media narratives, survey data, landmark research, and regulatory developments. Early anxieties (rooted in Cold War geopolitics and science fiction) focused on physical robots and autonomous weapons. In contrast, contemporary debates focus on algorithmic bias, misinformation, job displacement, and existential risks posed by advanced systems, such as Large Language Models (LLMs). This article examines the impact of key scholarly contributions, significant events, and regulatory milestones on public perception and governance approaches. Building on this context, this study proposes an improved LLM safety scoring system that prioritizes existential risk mitigation, transparency, and governance accountability. Applying the proposed framework to leading AI developers reveals significant variation in safety commitments. The results underscore how weighting choices affect rankings. Comparative analysis with existing indices highlights the importance of nuanced, multidimensional evaluation methods. The paper concludes by identifying pressing governance challenges, including the need for global cooperation, robust interpretability, and ongoing monitoring of harm in high-stakes domains. These findings demonstrate that AI safety is not static but somewhat shaped by historical context, technical capabilities, and societal values—requiring the continuous adaptation of both policy and evaluation frameworks to align AI systems with human interests.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"28 ","pages":"Article 200583"},"PeriodicalIF":4.3,"publicationDate":"2025-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145221296","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Federated learning for cyber attack detection to enhance security in protection schemes of cyber-physical energy systems 基于联邦学习的网络攻击检测，提高网络物理能源系统防护方案的安全性

IF 4.3

Intelligent Systems with Applications

Pub Date : 2025-09-25 DOI: 10.1016/j.iswa.2025.200590

Lei Du, Qingzhi Zhu

Cyber-attacks increasingly target the protection systems that safeguard cyber-physical energy systems (CPES), making it more difficult to deliver security and reliability requirements. The protection schemes in power grids, which depend on real-time forecasts from digital relays and Apple devices, require detection of physical faults and, simultaneously, malicious cyber attacks. This paper developed a decentralized federated learning-based framework to assist with the detection of cyber attacks in the protection schemes of cyber-physical energy systems (CPES), with the goals of privacy preservation and scalability. Attention was paid to the whole range of threats, including false data injection (FDI), man-in-the-middle, replay, and denial of service (DoS) across distributed substations without centralization of raw datasets. A lightweight neural network model was trained locally before being aggregated using federated averaging to develop a collaborative approach to learning across multiple substations. Based on the 3-machine, 9 bus case, simulations were run with synthetic attack datasets. The proposed method achieved an average detection accuracy of 96.7% while also preserving the confidentiality and non-disclosure of data. The study also highlighted some of the challenges related to implementation, conceptual drift, and the computational limits of hosting the solution, thereby providing a better understanding of planning and deploying the solution in smart grid applications.

网络攻击越来越多地针对保护网络物理能源系统（CPES）的保护系统，这使得满足安全性和可靠性要求变得更加困难。电网中的保护方案依赖于数字继电器和苹果设备的实时预测，需要检测物理故障，同时还要检测恶意网络攻击。本文开发了一个分散的基于学习的联邦框架，以帮助检测网络物理能源系统（CPES）保护方案中的网络攻击，目标是保护隐私和可扩展性。关注的是整个威胁范围，包括虚假数据注入（FDI），中间人，重播和拒绝服务（DoS）跨分布式变电站，没有集中的原始数据集。在使用联邦平均进行聚合之前，先对轻量级神经网络模型进行局部训练，以开发跨多个变电站学习的协作方法。在3机9总线的情况下，利用综合攻击数据集进行了仿真。该方法在保证数据保密性和非披露性的同时，平均检测准确率达到96.7%。该研究还强调了与实施、概念漂移和托管解决方案的计算限制相关的一些挑战，从而为在智能电网应用中规划和部署解决方案提供了更好的理解。

{"title":"Federated learning for cyber attack detection to enhance security in protection schemes of cyber-physical energy systems","authors":"Lei Du, Qingzhi Zhu","doi":"10.1016/j.iswa.2025.200590","DOIUrl":"10.1016/j.iswa.2025.200590","url":null,"abstract":"<div><div>Cyber-attacks increasingly target the protection systems that safeguard cyber-physical energy systems (CPES), making it more difficult to deliver security and reliability requirements. The protection schemes in power grids, which depend on real-time forecasts from digital relays and Apple devices, require detection of physical faults and, simultaneously, malicious cyber attacks. This paper developed a decentralized federated learning-based framework to assist with the detection of cyber attacks in the protection schemes of cyber-physical energy systems (CPES), with the goals of privacy preservation and scalability. Attention was paid to the whole range of threats, including false data injection (FDI), man-in-the-middle, replay, and denial of service (DoS) across distributed substations without centralization of raw datasets. A lightweight neural network model was trained locally before being aggregated using federated averaging to develop a collaborative approach to learning across multiple substations. Based on the 3-machine, 9 bus case, simulations were run with synthetic attack datasets. The proposed method achieved an average detection accuracy of 96.7% while also preserving the confidentiality and non-disclosure of data. The study also highlighted some of the challenges related to implementation, conceptual drift, and the computational limits of hosting the solution, thereby providing a better understanding of planning and deploying the solution in smart grid applications.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"28 ","pages":"Article 200590"},"PeriodicalIF":4.3,"publicationDate":"2025-09-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145221299","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Personalized privacy in OSNs: Evaluating deep learning models for context-aware image editing osn中的个性化隐私：评估上下文感知图像编辑的深度学习模型

IF 4.3

Intelligent Systems with Applications

Pub Date : 2025-09-25 DOI: 10.1016/j.iswa.2025.200581

Gelareh Hasel Mehri , Georgi Kostov , Bernardo Breve , Andrei Jalba , Nicola Zannone

Online Social Networks (OSNs) have become a cornerstone of digital interaction, enabling users to easily create and share content. While these platforms offer numerous benefits, they also expose users to privacy risks such as cyberstalking and identity theft. To address these concerns, OSNs typically provide access control mechanisms that allow users to regulate content visibility. However, these mechanisms often assume that content is managed by individual users and focus primarily on preserving content integrity, which may discourage users from sharing sensitive information. In this work, we propose a privacy model that empowers users to conceal sensitive content in images according to their preferences, expressed by means of policies. Our approach employs a multi-stage pipeline that includes segmentation for object localization, scene graphs and distance metrics for determining object ownership, and inpainting techniques for editing. We investigate the use of advanced deep learning models to implement the privacy model, aiming to provide personalized privacy controls while maintaining high image fidelity. To evaluate the proposed model, we conducted a user study with 20 participants. The user study highlights that ownership is the most significant factor influencing user perceptions of policy enforcement compliance, with less impact from localization and editing. The results also reveal that participants are generally willing to adopt the fully automated privacy model for selectively editing images in OSNs based on viewer identity, although some prefer alternative use cases, such as editing or censorship tools. Participants also raised concerns about the potential misuse of the model, supporting our choice of excluding an option for object replacement.

在线社交网络（OSNs）已经成为数字交互的基石，使用户能够轻松地创建和共享内容。虽然这些平台提供了许多好处，但它们也使用户面临网络跟踪和身份盗用等隐私风险。为了解决这些问题，osn通常提供允许用户调节内容可见性的访问控制机制。然而，这些机制通常假设内容是由单个用户管理的，并且主要关注于保持内容的完整性，这可能会阻碍用户共享敏感信息。在这项工作中，我们提出了一个隐私模型，使用户能够根据自己的偏好隐藏图像中的敏感内容，并通过策略来表达。我们的方法采用多阶段管道，包括用于对象定位的分割，用于确定对象所有权的场景图和距离度量，以及用于编辑的绘图技术。我们研究了使用先进的深度学习模型来实现隐私模型，旨在提供个性化的隐私控制，同时保持高图像保真度。为了评估所提出的模型，我们对20名参与者进行了用户研究。用户研究强调，所有权是影响用户对政策执行合规性看法的最重要因素，本地化和编辑的影响较小。研究结果还显示，参与者普遍愿意采用全自动隐私模型，根据观看者身份选择性地编辑osn中的图像，尽管有些人更喜欢其他用例，如编辑或审查工具。与会者还提出了对模型可能被滥用的担忧，支持我们排除对象替换选项的选择。

{"title":"Personalized privacy in OSNs: Evaluating deep learning models for context-aware image editing","authors":"Gelareh Hasel Mehri , Georgi Kostov , Bernardo Breve , Andrei Jalba , Nicola Zannone","doi":"10.1016/j.iswa.2025.200581","DOIUrl":"10.1016/j.iswa.2025.200581","url":null,"abstract":"<div><div>Online Social Networks (OSNs) have become a cornerstone of digital interaction, enabling users to easily create and share content. While these platforms offer numerous benefits, they also expose users to privacy risks such as cyberstalking and identity theft. To address these concerns, OSNs typically provide access control mechanisms that allow users to regulate content visibility. However, these mechanisms often assume that content is managed by individual users and focus primarily on preserving content integrity, which may discourage users from sharing sensitive information. In this work, we propose a privacy model that empowers users to conceal sensitive content in images according to their preferences, expressed by means of policies. Our approach employs a multi-stage pipeline that includes segmentation for object localization, scene graphs and distance metrics for determining object ownership, and inpainting techniques for editing. We investigate the use of advanced deep learning models to implement the privacy model, aiming to provide personalized privacy controls while maintaining high image fidelity. To evaluate the proposed model, we conducted a user study with 20 participants. The user study highlights that ownership is the most significant factor influencing user perceptions of policy enforcement compliance, with less impact from localization and editing. The results also reveal that participants are generally willing to adopt the fully automated privacy model for selectively editing images in OSNs based on viewer identity, although some prefer alternative use cases, such as editing or censorship tools. Participants also raised concerns about the potential misuse of the model, supporting our choice of excluding an option for object replacement.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"28 ","pages":"Article 200581"},"PeriodicalIF":4.3,"publicationDate":"2025-09-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145267361","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

New Harris Hawks algorithms for the Close-Enough Traveling Salesman Problem 足够近旅行商问题的新Harris Hawks算法

IF 4.3

Intelligent Systems with Applications

Pub Date : 2025-09-22 DOI: 10.1016/j.iswa.2025.200586

Tansel Dokeroglu, Deniz Canturk

This study introduces a novel application of the Harris Hawks Optimization (HHO) algorithm to the Close-Enough Traveling Salesman Problem (CETSP), a challenging combinatorial optimization problem where circular neighborhoods rather than exact coordinates represent target points. To tackle the CETSP’s spatial complexity and high-dimensional solution space, we develop new HHO algorithms, including a parallel multi-population variant designed using the OpenMP framework. This parallel algorithm allows multiple subpopulations to evolve simultaneously, increasing diversity and computational efficiency, particularly on large-scale and real-time instances. Furthermore, new problem-specific exploration and exploitation operators are introduced, tailored to the CETSP’s geometric structure, enabling better guidance of the search process toward high-quality solutions. A comprehensive empirical evaluation is conducted on 47 benchmark instances, encompassing synthetic problem instances and a real-world robotic welding scenario in automotive manufacturing. The results show that the proposed methods outperform existing state-of-the-art techniques such as Genetic Algorithm (GA), Memetic Algorithm (MA-CETSP) and Variable Neighborhood Search (VNS)-based approaches, achieving 18 new best-known solutions. The experimental findings underline the strong convergence behavior, robustness across diverse problem sizes, and practical applicability of the algorithm. Additionally, the algorithm’s modular and extensible structure leads the way for future adaptations to multi-objective and dynamic versions of CETSP, broadening its relevance for both academic research and industrial deployment.

本文介绍了Harris Hawks Optimization （HHO）算法在近距离旅行商问题（CETSP）中的一种新应用，这是一个具有挑战性的组合优化问题，其中圆形邻域而不是精确坐标表示目标点。为了解决CETSP的空间复杂性和高维解空间，我们开发了新的HHO算法，包括使用OpenMP框架设计的并行多种群变体。这种并行算法允许多个子种群同时进化，增加了多样性和计算效率，特别是在大规模和实时实例上。此外，针对CETSP的几何结构，还引入了新的针对特定问题的勘探和开发操作方法，从而更好地指导搜索过程，以获得高质量的解决方案。对47个基准实例进行了全面的实证评估，其中包括汽车制造中的综合问题实例和实际机器人焊接场景。结果表明，所提出的方法优于现有的最先进的技术，如遗传算法（GA），模因算法（MA-CETSP）和基于变量邻域搜索（VNS）的方法，实现了18个新的最知名的解决方案。实验结果强调了该算法的强收敛性、跨不同问题规模的鲁棒性和实际适用性。此外，该算法的模块化和可扩展结构为未来适应多目标和动态版本的CETSP铺平了道路，扩大了其在学术研究和工业部署方面的相关性。

{"title":"New Harris Hawks algorithms for the Close-Enough Traveling Salesman Problem","authors":"Tansel Dokeroglu, Deniz Canturk","doi":"10.1016/j.iswa.2025.200586","DOIUrl":"10.1016/j.iswa.2025.200586","url":null,"abstract":"<div><div>This study introduces a novel application of the Harris Hawks Optimization (HHO) algorithm to the Close-Enough Traveling Salesman Problem (CETSP), a challenging combinatorial optimization problem where circular neighborhoods rather than exact coordinates represent target points. To tackle the CETSP’s spatial complexity and high-dimensional solution space, we develop new HHO algorithms, including a parallel multi-population variant designed using the OpenMP framework. This parallel algorithm allows multiple subpopulations to evolve simultaneously, increasing diversity and computational efficiency, particularly on large-scale and real-time instances. Furthermore, new problem-specific exploration and exploitation operators are introduced, tailored to the CETSP’s geometric structure, enabling better guidance of the search process toward high-quality solutions. A comprehensive empirical evaluation is conducted on 47 benchmark instances, encompassing synthetic problem instances and a real-world robotic welding scenario in automotive manufacturing. The results show that the proposed methods outperform existing state-of-the-art techniques such as Genetic Algorithm (GA), Memetic Algorithm (MA-CETSP) and Variable Neighborhood Search (VNS)-based approaches, achieving 18 new best-known solutions. The experimental findings underline the strong convergence behavior, robustness across diverse problem sizes, and practical applicability of the algorithm. Additionally, the algorithm’s modular and extensible structure leads the way for future adaptations to multi-objective and dynamic versions of CETSP, broadening its relevance for both academic research and industrial deployment.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"28 ","pages":"Article 200586"},"PeriodicalIF":4.3,"publicationDate":"2025-09-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145158461","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Artificial intelligence-driven green innovation in packaging: A systematic review of adoption and diffusion challenges 包装中人工智能驱动的绿色创新：采用和扩散挑战的系统回顾

IF 4.3

Intelligent Systems with Applications

Pub Date : 2025-09-22 DOI: 10.1016/j.iswa.2025.200589

Ye Ma, Nor Hidayati Zakaria, Basheer Al-Haimi, Chen Wu

Global concern about environmental protection has intensified the demand for sustainable packaging solutions. Integrating artificial intelligence (AI) into green innovation offers a transformative way to address these challenges. This study applies a systematic literature review (SLR) guided by the PRISMA 2020 framework to examine recent AI-powered packaging innovations. Forty-eight peer-reviewed articles, published between 2020 and 2025, were analyzed. The findings show that Machine Learning, Deep Learning, and general AI applications are the most frequently adopted technologies. Biodegradable packaging materials and smart packaging systems represent the main sustainable packaging types. AI applications are concentrated in process optimization, smart packaging monitoring, fraud detection, computer vision, and natural language processing. However, widespread adoption faces obstacles such as high costs, technical complexity, and regulatory uncertainty. Future trends highlight the importance of scalable technologies, advanced AI models, integration with the circular economy, and interdisciplinary collaboration. This review provides a structured framework to guide academics, industry practitioners, and policymakers in adopting AI-driven green innovation for sustainable packaging.

全球对环境保护的关注加剧了对可持续包装解决方案的需求。将人工智能（AI）融入绿色创新为应对这些挑战提供了一种变革性的方式。本研究采用以PRISMA 2020框架为指导的系统文献综述（SLR）来研究最近的人工智能包装创新。研究人员分析了2020年至2025年间发表的48篇同行评议文章。研究结果表明，机器学习、深度学习和通用人工智能应用是最常采用的技术。可生物降解的包装材料和智能包装系统代表了主要的可持续包装类型。人工智能应用主要集中在流程优化、智能包装监控、欺诈检测、计算机视觉和自然语言处理等方面。然而，广泛采用面临着诸如高成本、技术复杂性和监管不确定性等障碍。未来的趋势强调了可扩展技术、先进的人工智能模型、与循环经济的整合以及跨学科合作的重要性。本综述提供了一个结构化的框架，以指导学者、行业从业者和政策制定者采用人工智能驱动的绿色创新来实现可持续包装。

{"title":"Artificial intelligence-driven green innovation in packaging: A systematic review of adoption and diffusion challenges","authors":"Ye Ma, Nor Hidayati Zakaria, Basheer Al-Haimi, Chen Wu","doi":"10.1016/j.iswa.2025.200589","DOIUrl":"10.1016/j.iswa.2025.200589","url":null,"abstract":"<div><div>Global concern about environmental protection has intensified the demand for sustainable packaging solutions. Integrating artificial intelligence (AI) into green innovation offers a transformative way to address these challenges. This study applies a systematic literature review (SLR) guided by the PRISMA 2020 framework to examine recent AI-powered packaging innovations. Forty-eight peer-reviewed articles, published between 2020 and 2025, were analyzed. The findings show that Machine Learning, Deep Learning, and general AI applications are the most frequently adopted technologies. Biodegradable packaging materials and smart packaging systems represent the main sustainable packaging types. AI applications are concentrated in process optimization, smart packaging monitoring, fraud detection, computer vision, and natural language processing. However, widespread adoption faces obstacles such as high costs, technical complexity, and regulatory uncertainty. Future trends highlight the importance of scalable technologies, advanced AI models, integration with the circular economy, and interdisciplinary collaboration. This review provides a structured framework to guide academics, industry practitioners, and policymakers in adopting AI-driven green innovation for sustainable packaging.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"28 ","pages":"Article 200589"},"PeriodicalIF":4.3,"publicationDate":"2025-09-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145221298","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

The semantic correlation mining method of multimodal data in constructing techno-economic knowledge graph of power grid 构建电网技术经济知识图谱中多模态数据的语义关联挖掘方法

IF 4.3

Intelligent Systems with Applications

Pub Date : 2025-09-21 DOI: 10.1016/j.iswa.2025.200588

Ling Qiu, Mengqi Pan, Nuoya Lv

Due to the diverse formats and complex structures of multimodal data, effectively managing its complexity and correlations remains challenging. Moreover, when dealing with large-scale data, traditional methods often encounter issues such as low computational efficiency and inaccurate results. This paper proposes a semantic association mining method for multimodal data. This method utilizes ETL technology to convert text and table data from different files into nodes and relational edges in the knowledge graph. By optimizing the word vector matrix through the skip character model, it can better capture the semantic information of text data and accurately reflect semantic similarity. Through integrating nodes such as equipment, design technologies and installation addresses, a technical and economic knowledge graph of the power grid is constructed. For the calculation of multimodal object associations, the data first undergoes label preprocessing, feature processing, and semantic relationship structuring before the association is computed using the cosine similarity formula. By using the association rule algorithm to mine the correlation relationships among time-series variables, potential correlations such as the operating status of equipment and the overall performance of the power grid can be discovered, thereby improving the understanding and prediction ability of the power grid’s operating status. The experimental results demonstrate that the proposed method achieves the highest accuracy and recall rate at 98.20 %, with an F-measure of 93.89 %, a bit error rate below 0.9, and a time consumption of approximately 7.34 s.

由于多模态数据的多种格式和复杂结构，有效管理其复杂性和相关性仍然具有挑战性。此外，在处理大规模数据时，传统方法往往会遇到计算效率低、结果不准确等问题。提出了一种多模态数据的语义关联挖掘方法。该方法利用ETL技术将不同文件中的文本和表格数据转换为知识图中的节点和关系边。通过跳过字符模型对词向量矩阵进行优化，可以更好地捕捉文本数据的语义信息，准确反映语义相似度。通过对设备、设计技术、安装地址等节点的整合，构建了电网的技术经济知识图谱。在计算多模态对象关联时，首先对数据进行标签预处理、特征处理和语义关系构建，然后使用余弦相似度公式计算关联。利用关联规则算法挖掘时间序列变量之间的相关关系，可以发现设备运行状态与电网整体性能等潜在的相关性，从而提高对电网运行状态的理解和预测能力。实验结果表明，该方法达到了98.20%的最高准确率和召回率，f值为93.89%，误码率低于0.9，时间消耗约为7.34 s。

{"title":"The semantic correlation mining method of multimodal data in constructing techno-economic knowledge graph of power grid","authors":"Ling Qiu, Mengqi Pan, Nuoya Lv","doi":"10.1016/j.iswa.2025.200588","DOIUrl":"10.1016/j.iswa.2025.200588","url":null,"abstract":"<div><div>Due to the diverse formats and complex structures of multimodal data, effectively managing its complexity and correlations remains challenging. Moreover, when dealing with large-scale data, traditional methods often encounter issues such as low computational efficiency and inaccurate results. This paper proposes a semantic association mining method for multimodal data. This method utilizes ETL technology to convert text and table data from different files into nodes and relational edges in the knowledge graph. By optimizing the word vector matrix through the skip character model, it can better capture the semantic information of text data and accurately reflect semantic similarity. Through integrating nodes such as equipment, design technologies and installation addresses, a technical and economic knowledge graph of the power grid is constructed. For the calculation of multimodal object associations, the data first undergoes label preprocessing, feature processing, and semantic relationship structuring before the association is computed using the cosine similarity formula. By using the association rule algorithm to mine the correlation relationships among time-series variables, potential correlations such as the operating status of equipment and the overall performance of the power grid can be discovered, thereby improving the understanding and prediction ability of the power grid’s operating status. The experimental results demonstrate that the proposed method achieves the highest accuracy and recall rate at 98.20 %, with an F-measure of 93.89 %, a bit error rate below 0.9, and a time consumption of approximately 7.34 s.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"28 ","pages":"Article 200588"},"PeriodicalIF":4.3,"publicationDate":"2025-09-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145221297","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A data-driven optimization approach for automated reviewer assignment using natural language processing 一种使用自然语言处理的数据驱动优化方法，用于自动审稿人分配

IF 4.3

Intelligent Systems with Applications

Pub Date : 2025-09-18 DOI: 10.1016/j.iswa.2025.200587

Meltem Aksoy , Seda Yanik , Mehmet Fatih Amasyali

In many settings, such as project or publication selection, expert reviewers play a pivotal role, as their assessments serve as the primary basis for determining a project's prospective value. The effectiveness of matching and assigning qualified experts to evaluate project proposals can substantially influence the quality of the selection process and, consequently, impact the funding organization's return on investment. Despite its importance, many funding organizations continue to rely on basic manual methods for assigning reviewers. This simplistic approach can compromise the quality of project selection and lead to suboptimal financial outcomes. Moreover, it may hinder the equitable distribution of review workloads and increase conflicts of interest between reviewers and applicants. Consequently, there is a pressing need for a systematic and automated method to enhance the reviewer assignment process.

In this study, we propose an optimization-based approach using natural language processing to automate the reviewer assignment process for project proposals. The proposed approach follows a structured three-stage methodology. First, a comprehensive database is constructed by collecting multilingual data on both proposals and reviewers. Second, word embedding techniques are used to represent texts as vectors, enabling the use of cosine similarity to quantify the relevance between each proposal and reviewer. Reviewer expertise and past evaluation performance are also analyzed using predefined knowledge rules. In the final stage, a multi-objective integer linear programming model assigns reviewers by optimizing proposal-reviewer similarity and reviewer competency while preventing conflicts of interest. Additionally, a max-min strategy is employed to ensure fair treatment of less-advantaged proposals, and two supplementary models are introduced to balance reviewer workloads. Experimental results on a real-world dataset from a regional development agency demonstrate that the proposed system significantly outperforms traditional manual assignment methods. We show that automated reviewer assignment prevents subjective judgements, together with reductions in time and cost of the assignment process.

在许多情况下，例如项目或出版物选择，专家审稿人扮演着关键的角色，因为他们的评估是确定项目预期价值的主要基础。匹配和分配合格专家来评估项目提案的有效性可以极大地影响选择过程的质量，从而影响资助组织的投资回报。尽管它很重要，但许多资助组织仍然依赖于基本的手工方法来分配审稿人。这种简单的方法可能会损害项目选择的质量，并导致次优的财务结果。此外，它可能阻碍审查工作量的公平分配，并增加审查者和申请人之间的利益冲突。因此，迫切需要一种系统和自动化的方法来增强审稿人分配过程。在这项研究中，我们提出了一种基于优化的方法，使用自然语言处理来自动化项目提案的审稿人分配过程。拟议的方法遵循结构化的三阶段方法。首先，通过收集提案和审稿人的多语种数据，构建一个全面的数据库。其次，使用词嵌入技术将文本表示为向量，从而可以使用余弦相似度来量化每个提案和审稿人之间的相关性。使用预定义的知识规则分析审稿人的专业知识和过去的评估绩效。最后，在避免利益冲突的同时，通过优化提案-审稿人相似性和审稿人能力，建立多目标整数线性规划模型分配审稿人。此外，采用了最大最小策略来确保公平对待劣势提案，并引入了两个补充模型来平衡审稿人的工作量。在一个区域发展机构的真实数据集上的实验结果表明，该系统显著优于传统的人工分配方法。我们展示了自动审稿人分配防止了主观判断，同时减少了分配过程的时间和成本。

{"title":"A data-driven optimization approach for automated reviewer assignment using natural language processing","authors":"Meltem Aksoy , Seda Yanik , Mehmet Fatih Amasyali","doi":"10.1016/j.iswa.2025.200587","DOIUrl":"10.1016/j.iswa.2025.200587","url":null,"abstract":"<div><div>In many settings, such as project or publication selection, expert reviewers play a pivotal role, as their assessments serve as the primary basis for determining a project's prospective value. The effectiveness of matching and assigning qualified experts to evaluate project proposals can substantially influence the quality of the selection process and, consequently, impact the funding organization's return on investment. Despite its importance, many funding organizations continue to rely on basic manual methods for assigning reviewers. This simplistic approach can compromise the quality of project selection and lead to suboptimal financial outcomes. Moreover, it may hinder the equitable distribution of review workloads and increase conflicts of interest between reviewers and applicants. Consequently, there is a pressing need for a systematic and automated method to enhance the reviewer assignment process.</div><div>In this study, we propose an optimization-based approach using natural language processing to automate the reviewer assignment process for project proposals. The proposed approach follows a structured three-stage methodology. First, a comprehensive database is constructed by collecting multilingual data on both proposals and reviewers. Second, word embedding techniques are used to represent texts as vectors, enabling the use of cosine similarity to quantify the relevance between each proposal and reviewer. Reviewer expertise and past evaluation performance are also analyzed using predefined knowledge rules. In the final stage, a multi-objective integer linear programming model assigns reviewers by optimizing proposal-reviewer similarity and reviewer competency while preventing conflicts of interest. Additionally, a max-min strategy is employed to ensure fair treatment of less-advantaged proposals, and two supplementary models are introduced to balance reviewer workloads. Experimental results on a real-world dataset from a regional development agency demonstrate that the proposed system significantly outperforms traditional manual assignment methods. We show that automated reviewer assignment prevents subjective judgements, together with reductions in time and cost of the assignment process.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"28 ","pages":"Article 200587"},"PeriodicalIF":4.3,"publicationDate":"2025-09-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145159057","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0