Pub Date : 2025-09-17DOI: 10.1016/j.iswa.2025.200582
Zander Wessels , Andries Engelbrecht
A novel approach to portfolio optimization is introduced using a variant of set-based particle swarm optimization (SBPSO), building upon the foundational work of Erwin and Engelbrecht. Although their contributions advanced the application of SBPSO to financial markets, this research addresses key practical challenges, specifically enhancing the treatment of covariance and expected returns and refining constraint implementations to align with real-world applications. Beyond algorithmic improvements, this article emphasizes the importance of robust evaluation methodologies and highlights the limitations of traditional backtesting frameworks, which often yield overly optimistic results. To overcome these biases, the study introduces a comprehensive simulation platform that mitigates issues such as survivorship and forward-looking bias. This provides a realistic assessment of the modified SBPSO’s financial performance under varying market conditions. The findings shift the focus from computational efficiency to the practical outcomes of profitability that are most relevant to investors.
{"title":"Enhanced set-based particle swarm optimization for portfolio management in a walk-forward paradigm","authors":"Zander Wessels , Andries Engelbrecht","doi":"10.1016/j.iswa.2025.200582","DOIUrl":"10.1016/j.iswa.2025.200582","url":null,"abstract":"<div><div>A novel approach to portfolio optimization is introduced using a variant of set-based particle swarm optimization (SBPSO), building upon the foundational work of Erwin and Engelbrecht. Although their contributions advanced the application of SBPSO to financial markets, this research addresses key practical challenges, specifically enhancing the treatment of covariance and expected returns and refining constraint implementations to align with real-world applications. Beyond algorithmic improvements, this article emphasizes the importance of robust evaluation methodologies and highlights the limitations of traditional backtesting frameworks, which often yield overly optimistic results. To overcome these biases, the study introduces a comprehensive simulation platform that mitigates issues such as survivorship and forward-looking bias. This provides a realistic assessment of the modified SBPSO’s financial performance under varying market conditions. The findings shift the focus from computational efficiency to the practical outcomes of profitability that are most relevant to investors.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"28 ","pages":"Article 200582"},"PeriodicalIF":4.3,"publicationDate":"2025-09-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145097476","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}
Pub Date : 2025-09-15DOI: 10.1016/j.iswa.2025.200584
Sinethemba H. Yakobi, Uchechukwu U. Nwodo
Vaccine instability contributes to the loss of up to 25 % of doses globally, a challenge intensified by the complexity of next-generation platforms such as mRNA–lipid nanoparticles (mRNA–LNPs), viral vectors, and protein subunits. Current regulatory frameworks (ICH Q5C, WHO TRS 1010) rely on static protocols that overlook platform-specific degradation mechanisms and real-world cold-chain variability. We introduce the Systems Biology–guided AI (SBg-AI) framework, a predictive stability platform integrating omics-derived biomarkers, real-time telemetry, and explainable machine learning. Leveraging recurrent and graph neural networks with Bayesian inference, SBg-AI forecasts degradation events with 89 % accuracy—validated in African and Southeast Asian supply chains. Federated learning ensures multi-manufacturer collaboration while preserving data privacy. In field trials, dynamic expiry predictions reduced mRNA vaccine wastage by 22 %. A phased regulatory roadmap supports transition from hybrid AI-empirical models (2024) to full AI-based stability determinations by 2030. By integrating mechanistic degradation science with real-time telemetry and regulatory-compliant AI, the SBg-AI framework transforms vaccine stability from retrospective batch testing to proactive, precision-guided assurance.
疫苗的不稳定性导致全球高达25%的剂量损失，下一代平台（如mrna -脂质纳米颗粒（mRNA-LNPs））、病毒载体和蛋白质亚基)的复杂性加剧了这一挑战。目前的监管框架（ICH Q5C, WHO TRS 1010）依赖于静态协议，忽略了平台特定的降解机制和现实世界的冷链可变性。我们介绍了系统生物学引导的人工智能（SBg-AI）框架，这是一个集成了组学衍生生物标志物、实时遥测和可解释机器学习的预测稳定性平台。利用贝叶斯推理的循环神经网络和图神经网络，SBg-AI预测退化事件的准确率为89%，在非洲和东南亚的供应链中得到了验证。联邦学习确保多制造商协作，同时保护数据隐私。在田间试验中，动态过期预测使mRNA疫苗的浪费减少了22%。分阶段的监管路线图支持从混合人工智能经验模型（2024年）过渡到2030年完全基于人工智能的稳定性确定。通过将机械降解科学与实时遥测和符合法规的人工智能相结合，SBg-AI框架将疫苗稳定性从回顾性批量检测转变为主动、精确指导的保证。
{"title":"AI-predictive vaccine stability: a systems biology framework to modernize regulatory testing and cold chain equity","authors":"Sinethemba H. Yakobi, Uchechukwu U. Nwodo","doi":"10.1016/j.iswa.2025.200584","DOIUrl":"10.1016/j.iswa.2025.200584","url":null,"abstract":"<div><div>Vaccine instability contributes to the loss of up to 25 % of doses globally, a challenge intensified by the complexity of next-generation platforms such as mRNA–lipid nanoparticles (mRNA–LNPs), viral vectors, and protein subunits. Current regulatory frameworks (ICH Q5C, WHO TRS 1010) rely on static protocols that overlook platform-specific degradation mechanisms and real-world cold-chain variability. We introduce the Systems Biology–guided AI (SBg-AI) framework, a predictive stability platform integrating omics-derived biomarkers, real-time telemetry, and explainable machine learning. Leveraging recurrent and graph neural networks with Bayesian inference, SBg-AI forecasts degradation events with 89 % accuracy—validated in African and Southeast Asian supply chains. Federated learning ensures multi-manufacturer collaboration while preserving data privacy. In field trials, dynamic expiry predictions reduced mRNA vaccine wastage by 22 %. A phased regulatory roadmap supports transition from hybrid AI-empirical models (2024) to full AI-based stability determinations by 2030. By integrating mechanistic degradation science with real-time telemetry and regulatory-compliant AI, the SBg-AI framework transforms vaccine stability from retrospective batch testing to proactive, precision-guided assurance.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"28 ","pages":"Article 200584"},"PeriodicalIF":4.3,"publicationDate":"2025-09-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145097475","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}
Pub Date : 2025-09-15DOI: 10.1016/j.iswa.2025.200580
Nuno Bento, Joana Rebelo, Marília Barandas
Robust and accurate evaluation metrics are crucial to test generative models and ensure their practical utility. However, the most common metrics heavily rely on the selected data representation and may not be strongly correlated with the ground truth, which itself can be difficult to obtain. This paper attempts to simplify this process by proposing a benchmark to compare data representations in an automatic manner, i.e. without relying on human evaluators. This is achieved through a simple test based on the assumption that samples with higher quality should lead to improved metric scores. Furthermore, we apply this benchmark on small, low-resolution image datasets to explore various representations, including embeddings finetuned either on the same dataset or on different datasets. An extensive evaluation shows the superiority of pretrained embeddings over randomly initialized representations, as well as evidence that embeddings trained on external, more diverse datasets outperform task-specific ones.
{"title":"Benchmarking deep neural representations for synthetic data evaluation","authors":"Nuno Bento, Joana Rebelo, Marília Barandas","doi":"10.1016/j.iswa.2025.200580","DOIUrl":"10.1016/j.iswa.2025.200580","url":null,"abstract":"<div><div>Robust and accurate evaluation metrics are crucial to test generative models and ensure their practical utility. However, the most common metrics heavily rely on the selected data representation and may not be strongly correlated with the ground truth, which itself can be difficult to obtain. This paper attempts to simplify this process by proposing a benchmark to compare data representations in an automatic manner, i.e. without relying on human evaluators. This is achieved through a simple test based on the assumption that samples with higher quality should lead to improved metric scores. Furthermore, we apply this benchmark on small, low-resolution image datasets to explore various representations, including embeddings finetuned either on the same dataset or on different datasets. An extensive evaluation shows the superiority of pretrained embeddings over randomly initialized representations, as well as evidence that embeddings trained on external, more diverse datasets outperform task-specific ones.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"28 ","pages":"Article 200580"},"PeriodicalIF":4.3,"publicationDate":"2025-09-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145097474","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}
As the Internet of Things expands, managing intelligent tasks in dynamic and heterogeneous environments has emerged as a primary challenge for processing-based systems at the network’s edge. A critical issue in this domain is the optimal allocation of tasks. A review of prior studies indicates that many existing approaches either focus on a single objective or suffer from instability and overestimation of decision values during the learning phase. This paper aims to bridge this by proposing an approach that utilizes reinforcement learning with a double Q-learning algorithm and a multi-objective reward function. Furthermore, the designed reward function facilitates intelligent decision-making under more realistic conditions by incorporating three essential factors: task execution delay, energy consumption of edge nodes, and computational load balancing across the nodes. The inputs for the proposed method encompass information such as task sizes, deadlines for each task, remaining energy in the nodes, computational power of the nodes, proximity to the edge nodes, and the current workload of each node. The method's output at any given moment is the decision regarding assigning any task to the most suitable node. Simulation results in a dynamic environment demonstrate that the proposed method outperforms traditional reinforcement learning algorithms. Specifically, the average task execution delay has been reduced by up to 23%, the energy consumption of the nodes has decreased by up to 18%, and load balancing among nodes has improved by up to 27%.
{"title":"Optimizing the distribution of tasks in Internet of Things using edge processing-based reinforcement learning","authors":"Mohsen Latifi, Nahideh Derakhshanfard, Hossein Heydari","doi":"10.1016/j.iswa.2025.200585","DOIUrl":"10.1016/j.iswa.2025.200585","url":null,"abstract":"<div><div>As the Internet of Things expands, managing intelligent tasks in dynamic and heterogeneous environments has emerged as a primary challenge for processing-based systems at the network’s edge. A critical issue in this domain is the optimal allocation of tasks. A review of prior studies indicates that many existing approaches either focus on a single objective or suffer from instability and overestimation of decision values during the learning phase. This paper aims to bridge this by proposing an approach that utilizes reinforcement learning with a double Q-learning algorithm and a multi-objective reward function. Furthermore, the designed reward function facilitates intelligent decision-making under more realistic conditions by incorporating three essential factors: task execution delay, energy consumption of edge nodes, and computational load balancing across the nodes. The inputs for the proposed method encompass information such as task sizes, deadlines for each task, remaining energy in the nodes, computational power of the nodes, proximity to the edge nodes, and the current workload of each node. The method's output at any given moment is the decision regarding assigning any task to the most suitable node. Simulation results in a dynamic environment demonstrate that the proposed method outperforms traditional reinforcement learning algorithms. Specifically, the average task execution delay has been reduced by up to 23%, the energy consumption of the nodes has decreased by up to 18%, and load balancing among nodes has improved by up to 27%.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"28 ","pages":"Article 200585"},"PeriodicalIF":4.3,"publicationDate":"2025-09-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145097473","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}
Pub Date : 2025-09-11DOI: 10.1016/j.iswa.2025.200578
Kaavya Rekanar , Martin J. Hayes , Ciarán Eising
Visual Question Answering (VQA) models serve a critical role in interpreting visual data and responding to textual queries, particularly within the domain of autonomous driving. These models enhance situational awareness and enable naturalistic interaction between passengers and vehicle systems. However, existing VQA architectures often underperform in driving contexts due to their generic design and lack of alignment with domain-specific perceptual cues. This study introduces a targeted enhancement strategy based on the integration of human visual attention patterns into VQA systems. The proposed approach investigates visual subjectivity by analysing human responses and gaze behaviours captured through an eye-tracking experiment conducted in a realistic driving scenario. This method enables the direct observation of authentic attention patterns and mitigates the limitations introduced by subjective self-reporting. From these findings, a Human Attention Filter (HAF) is constructed to selectively preserve task-relevant features while suppressing visually distracting but semantically irrelevant content. Three VQA models – LXMERT, ViLBERT, and ViLT – are evaluated to demonstrate the adaptability and impact of HAF across different visual representation strategies, including region-based and patch-based architectures. Case studies involving LXMERT and ViLBERT are conducted to assess the integration of the HAF within region-based multimodal pipelines, showing measurable improvements in performance and alignment with human-like attention. Quantitative analysis reveals statistically significant performance trends correlated with driving experience, highlighting cognitive variability among human participants and informing model interpretability. In addition, failure cases are examined to identify potential limitations introduced by attention filtering, offering critical insight into the boundaries of gaze-guided model alignment.The findings validate the effectiveness of human-informed filtering for improving both accuracy and transparency in autonomous VQA systems, and present HAF as a sustainable, cognitively aligned strategy for advancing trustworthy AI in real-world environments.
{"title":"Mimicking human attention in driving scenarios for enhanced Visual Question Answering: Insights from eye-tracking and the human attention filter","authors":"Kaavya Rekanar , Martin J. Hayes , Ciarán Eising","doi":"10.1016/j.iswa.2025.200578","DOIUrl":"10.1016/j.iswa.2025.200578","url":null,"abstract":"<div><div>Visual Question Answering (VQA) models serve a critical role in interpreting visual data and responding to textual queries, particularly within the domain of autonomous driving. These models enhance situational awareness and enable naturalistic interaction between passengers and vehicle systems. However, existing VQA architectures often underperform in driving contexts due to their generic design and lack of alignment with domain-specific perceptual cues. This study introduces a targeted enhancement strategy based on the integration of human visual attention patterns into VQA systems. The proposed approach investigates visual subjectivity by analysing human responses and gaze behaviours captured through an eye-tracking experiment conducted in a realistic driving scenario. This method enables the direct observation of authentic attention patterns and mitigates the limitations introduced by subjective self-reporting. From these findings, a Human Attention Filter (HAF) is constructed to selectively preserve task-relevant features while suppressing visually distracting but semantically irrelevant content. Three VQA models – LXMERT, ViLBERT, and ViLT – are evaluated to demonstrate the adaptability and impact of HAF across different visual representation strategies, including region-based and patch-based architectures. Case studies involving LXMERT and ViLBERT are conducted to assess the integration of the HAF within region-based multimodal pipelines, showing measurable improvements in performance and alignment with human-like attention. Quantitative analysis reveals statistically significant performance trends correlated with driving experience, highlighting cognitive variability among human participants and informing model interpretability. In addition, failure cases are examined to identify potential limitations introduced by attention filtering, offering critical insight into the boundaries of gaze-guided model alignment.The findings validate the effectiveness of human-informed filtering for improving both accuracy and transparency in autonomous VQA systems, and present HAF as a sustainable, cognitively aligned strategy for advancing trustworthy AI in real-world environments.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"28 ","pages":"Article 200578"},"PeriodicalIF":4.3,"publicationDate":"2025-09-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145061040","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}
Accurate long-term prediction in industrial processes is essential for efficient control and operation. This study investigates the use of artificial neural networks (ANNs) for forecasting temperature in complex thermal systems, with a focus on enhancing model robustness under real-world conditions. A key innovation in this work is the intentional introduction of Gaussian noise into the training data to emulate sensor inaccuracies and environmental uncertainties, thereby improving the network's generalization capability. The target application is the prediction of water temperature in a non-stirred reservoir heated by two electric heaters, where phase change, thermal gradients, and sensor placement introduce significant modeling challenges. The proposed feedforward neural network architecture, comprising 90 neurons across three hidden layers, demonstrated a substantial reduction in long-term prediction error from 11.23 % to 2.02 % when trained with noise-augmented data. This result highlights the effectiveness of noise injection as a regularization strategy for improving performance in forecasting tasks. The study further contrasts this approach with Random Forest model and confirms the superior generalization and stability of the noise-trained ANN. These findings establish a scalable methodology for improving predictive accuracy in industrial systems characterized by limited data, strong nonlinearities, and uncertain measurements.
{"title":"Improving long-term prediction in industrial processes using neural networks with noise-added training data","authors":"Mohammadhossein Ghadimi Mahanipoor , Amirhossein Fathi","doi":"10.1016/j.iswa.2025.200579","DOIUrl":"10.1016/j.iswa.2025.200579","url":null,"abstract":"<div><div>Accurate long-term prediction in industrial processes is essential for efficient control and operation. This study investigates the use of artificial neural networks (ANNs) for forecasting temperature in complex thermal systems, with a focus on enhancing model robustness under real-world conditions. A key innovation in this work is the intentional introduction of Gaussian noise into the training data to emulate sensor inaccuracies and environmental uncertainties, thereby improving the network's generalization capability. The target application is the prediction of water temperature in a non-stirred reservoir heated by two electric heaters, where phase change, thermal gradients, and sensor placement introduce significant modeling challenges. The proposed feedforward neural network architecture, comprising 90 neurons across three hidden layers, demonstrated a substantial reduction in long-term prediction error from 11.23 % to 2.02 % when trained with noise-augmented data. This result highlights the effectiveness of noise injection as a regularization strategy for improving performance in forecasting tasks. The study further contrasts this approach with Random Forest model and confirms the superior generalization and stability of the noise-trained ANN. These findings establish a scalable methodology for improving predictive accuracy in industrial systems characterized by limited data, strong nonlinearities, and uncertain measurements.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"28 ","pages":"Article 200579"},"PeriodicalIF":4.3,"publicationDate":"2025-09-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145050158","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}
Pub Date : 2025-09-07DOI: 10.1016/j.iswa.2025.200576
Ivo Façoco, Rafaela Carvalho, Luís Rosado
The rapid advancement of both wafer manufacturing and AI technologies is reshaping the semiconductor industry. As chip features become smaller and more intricate, the variety and complexity of defects continue to grow, making defect detection increasingly challenging. Meanwhile, AI has made significant strides in unsupervised anomaly detection and supervised defect segmentation, yet its application to wafer inspection remains underexplored. This work bridges these fields by investigating cutting-edge lightweight AI techniques for automated inspection of current generation of silicon wafers. Our study leverages a newly curated dataset comprising 1,055 images of 300 mm wafers, annotated with 6,861 defect labels across seven distinct types, along with PASS/FAIL decisions. From a data-centric perspective, we introduce a novel unsupervised dataset-splitting approach to ensure balanced representation of defect classes and image features. Using the DINO-ViT-S/8 model for feature extraction, our method achieves 96% coverage while maintaining the target 20% test ratio for both individual defects and PASS/FAIL classification. From a model-centric perspective, we benchmark several recent methods for unsupervised anomaly detection and supervised defect segmentation. For unsupervised anomaly detection, EfficientAD obtains the best performance for both pixel-level and image-wise metrics, with F1-scores of 75.14% and 82.35%, respectively. For supervised defect segmentation, UPerNet-Swin achieves the highest performance, with a pixel-level mDice of 47.90 and a mask-level F1-score of 57.45. To facilitate deployment in high-throughput conditions, we conduct a comparative analysis of computational efficiency. Finally, we explore a dual-stage output fusion approach that integrates the best-performing unsupervised anomaly detection and supervised segmentation models to refine PASS/FAIL decisions by incorporating defect severity.
晶圆制造和人工智能技术的快速发展正在重塑半导体产业。随着芯片特征越来越小、越来越复杂，缺陷的种类和复杂性也在不断增加，使得缺陷检测越来越具有挑战性。与此同时，人工智能在无监督异常检测和监督缺陷分割方面取得了重大进展，但其在晶圆检测中的应用仍未得到充分探索。这项工作通过研究用于当前一代硅片自动检测的尖端轻量级人工智能技术，将这些领域联系起来。我们的研究利用了一个新整理的数据集，其中包括1055张300毫米晶圆的图像，标注了7种不同类型的6861个缺陷标签，以及通过/不通过的决定。从以数据为中心的角度来看，我们引入了一种新的无监督数据集分割方法，以确保缺陷类和图像特征的平衡表示。使用dino - viti - s /8模型进行特征提取，我们的方法实现了96%的覆盖率，同时对单个缺陷和PASS/FAIL分类保持20%的目标测试比率。从以模型为中心的角度来看，我们对几种最新的无监督异常检测和监督缺陷分割方法进行了基准测试。对于无监督异常检测，EfficientAD在像素级和图像级指标上都获得了最佳性能，f1得分分别为75.14%和82.35%。对于监督缺陷分割，supernet - swin达到了最高的性能，像素级的mdevice为47.90，掩码级的F1-score为57.45。为了便于在高吞吐量条件下部署，我们对计算效率进行了比较分析。最后，我们探索了一种双阶段输出融合方法，该方法集成了性能最好的无监督异常检测和监督分割模型，通过结合缺陷严重程度来改进PASS/FAIL决策。
{"title":"Towards efficient wafer visual inspection: Exploring novel lightweight approaches for anomaly detection and defect segmentation","authors":"Ivo Façoco, Rafaela Carvalho, Luís Rosado","doi":"10.1016/j.iswa.2025.200576","DOIUrl":"10.1016/j.iswa.2025.200576","url":null,"abstract":"<div><div>The rapid advancement of both wafer manufacturing and AI technologies is reshaping the semiconductor industry. As chip features become smaller and more intricate, the variety and complexity of defects continue to grow, making defect detection increasingly challenging. Meanwhile, AI has made significant strides in unsupervised anomaly detection and supervised defect segmentation, yet its application to wafer inspection remains underexplored. This work bridges these fields by investigating cutting-edge lightweight AI techniques for automated inspection of current generation of silicon wafers. Our study leverages a newly curated dataset comprising 1,055 images of 300 mm wafers, annotated with 6,861 defect labels across seven distinct types, along with PASS/FAIL decisions. From a data-centric perspective, we introduce a novel unsupervised dataset-splitting approach to ensure balanced representation of defect classes and image features. Using the DINO-ViT-S/8 model for feature extraction, our method achieves 96% coverage while maintaining the target 20% test ratio for both individual defects and PASS/FAIL classification. From a model-centric perspective, we benchmark several recent methods for unsupervised anomaly detection and supervised defect segmentation. For unsupervised anomaly detection, EfficientAD obtains the best performance for both pixel-level and image-wise metrics, with F1-scores of 75.14% and 82.35%, respectively. For supervised defect segmentation, UPerNet-Swin achieves the highest performance, with a pixel-level mDice of 47.90 and a mask-level F1-score of 57.45. To facilitate deployment in high-throughput conditions, we conduct a comparative analysis of computational efficiency. Finally, we explore a dual-stage output fusion approach that integrates the best-performing unsupervised anomaly detection and supervised segmentation models to refine PASS/FAIL decisions by incorporating defect severity.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"28 ","pages":"Article 200576"},"PeriodicalIF":4.3,"publicationDate":"2025-09-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145021018","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}
Pub Date : 2025-09-02DOI: 10.1016/j.iswa.2025.200573
Genjuan Ma, Yan Li
Early detection of liver cirrhosis remains problematic due to its asymptomatic onset and the inherent class imbalance in clinical data. This study conducts a comprehensive evaluation of machine learning models for predicting cirrhosis stages, with a focus on addressing these challenges. An approach employing Quadratic Discriminant Analysis (QDA) is benchmarked against seven other models, including powerful ensembles like Stacking and HistGradientBoosting, on a clinical dataset. Methodologies such as SMOTE oversampling, stratified data splitting, and class-specific covariance estimation were implemented to manage data complexity. The results demonstrate that a Stacking ensemble achieves the highest overall predictive performance with a micro-AUC of 0.80. The proposed QDA method also proves to be a highly effective and competitive model, achieving a robust AUC of 0.76 and outperforming several specialized imbalance-learning algorithms. Crucially, QDA offers this strong performance with exceptional computational efficiency. These findings show that while complex ensembles can yield top-tier accuracy, QDA’s capacity to model non-linear feature associations makes it a powerful and practical choice for the diagnosis of cirrhosis.
{"title":"Liver cirrhosis prediction: The employment of the machine learning-based approaches","authors":"Genjuan Ma, Yan Li","doi":"10.1016/j.iswa.2025.200573","DOIUrl":"10.1016/j.iswa.2025.200573","url":null,"abstract":"<div><div>Early detection of liver cirrhosis remains problematic due to its asymptomatic onset and the inherent class imbalance in clinical data. This study conducts a comprehensive evaluation of machine learning models for predicting cirrhosis stages, with a focus on addressing these challenges. An approach employing Quadratic Discriminant Analysis (QDA) is benchmarked against seven other models, including powerful ensembles like Stacking and HistGradientBoosting, on a clinical dataset. Methodologies such as SMOTE oversampling, stratified data splitting, and class-specific covariance estimation were implemented to manage data complexity. The results demonstrate that a Stacking ensemble achieves the highest overall predictive performance with a micro-AUC of 0.80. The proposed QDA method also proves to be a highly effective and competitive model, achieving a robust AUC of 0.76 and outperforming several specialized imbalance-learning algorithms. Crucially, QDA offers this strong performance with exceptional computational efficiency. These findings show that while complex ensembles can yield top-tier accuracy, QDA’s capacity to model non-linear feature associations makes it a powerful and practical choice for the diagnosis of cirrhosis.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"28 ","pages":"Article 200573"},"PeriodicalIF":4.3,"publicationDate":"2025-09-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145097472","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}
Pub Date : 2025-09-01DOI: 10.1016/j.iswa.2025.200575
Ramen Ghosh
We develop a data-driven framework for long-term forecasting of stochastic dynamics on evolving networked infrastructure systems using neural approximations of Koopman operators. In real-world nonlinear systems, the exact Koopman operator is infinite-dimensional and generally unavailable in closed form, necessitating learned finite-dimensional surrogates. Focusing on applications such as traffic flow and power grid oscillations, we model the underlying dynamics as random graph-driven nonlinear processes and introduce a graph-informed neural architecture that learns approximate Koopman eigenfunctions to capture system evolution over time. Our key contribution is the joint treatment of stochastic network evolution, Koopman operator learning, and phase-transition-induced breakdowns in forecasting. We identify critical regimes—arising from graph connectivity shifts or load-induced bifurcations—where the effective forecasting horizon collapses due to spectral degeneracy in the learned Koopman operator. We establish sufficient conditions under which this collapse occurs and propose regularization techniques to mitigate representational breakdown. Numerical experiments on traffic and power networks validate the proposed method and confirm the emergence of critical behavior. These results not only highlight the challenges of forecasting near structural transitions, but also suggest that spectral collapse may serve as a diagnostic signal for detecting phase transitions in dynamic networks. Our contributions unify spectral operator theory, random dynamical systems, and neural forecasting into a control-theoretic framework for real-time intelligent infrastructure. To our knowledge, this is the first work to jointly study Koopman operator learning, stochastic network evolution, and forecasting collapse induced by graph-theoretic phase transitions.
{"title":"Neural Koopman forecasting for critical transitions in infrastructure networks","authors":"Ramen Ghosh","doi":"10.1016/j.iswa.2025.200575","DOIUrl":"10.1016/j.iswa.2025.200575","url":null,"abstract":"<div><div>We develop a data-driven framework for long-term forecasting of stochastic dynamics on evolving networked infrastructure systems using neural approximations of Koopman operators. In real-world nonlinear systems, the exact Koopman operator is infinite-dimensional and generally unavailable in closed form, necessitating learned finite-dimensional surrogates. Focusing on applications such as traffic flow and power grid oscillations, we model the underlying dynamics as random graph-driven nonlinear processes and introduce a graph-informed neural architecture that learns approximate Koopman eigenfunctions to capture system evolution over time. Our key contribution is the joint treatment of stochastic network evolution, Koopman operator learning, and phase-transition-induced breakdowns in forecasting. We identify critical regimes—arising from graph connectivity shifts or load-induced bifurcations—where the effective forecasting horizon collapses due to spectral degeneracy in the learned Koopman operator. We establish sufficient conditions under which this collapse occurs and propose regularization techniques to mitigate representational breakdown. Numerical experiments on traffic and power networks validate the proposed method and confirm the emergence of critical behavior. These results not only highlight the challenges of forecasting near structural transitions, but also suggest that spectral collapse may serve as a diagnostic signal for detecting phase transitions in dynamic networks. Our contributions unify spectral operator theory, random dynamical systems, and neural forecasting into a control-theoretic framework for real-time intelligent infrastructure. To our knowledge, this is the first work to jointly study Koopman operator learning, stochastic network evolution, and forecasting collapse induced by graph-theoretic phase transitions.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"27 ","pages":"Article 200575"},"PeriodicalIF":4.3,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144921887","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}
Tree-based ensemble methods, such as Extreme Gradient Boosting (XGBoost) and Gradient Boosting models (GBM), are widely used for supervised learning due to their strong predictive capabilities. However, their complex architectures often hinder interpretability. This paper extends a lattice-theoretic framework originally developed for Random Forests to boosting algorithms, enabling a structured analysis of their internal logic via formal concept analysis (FCA).
We formally adapt four conceptual views: leaf, tree, tree predicate, and interordinal predicate to account for the sequential learning and optimization processes unique to boosting. Using the binary-class version of the car evaluation dataset from the OpenML CC18 benchmark suite, we conduct a systematic parameter study to examine how hyperparameters, such as tree depth and the number of trees, affect both model performance and conceptual complexity. Random Forest results from prior literature are used as a comparative baseline.
The results show that XGBoost yields the highest test accuracy, while GBM demonstrates greater stability in generalization error. Conceptually, boosting methods generate more compact and interpretable leaf views but preserve rich structural information in higher-level views. In contrast, Random Forests tend to produce denser and more redundant concept lattices. These trade-offs highlight how boosting methods, when interpreted through FCA, can strike a balance between performance and transparency.
Overall, this work contributes to explainable AI by demonstrating how lattice-based conceptual views can be systematically extended to complex boosting models, offering interpretable insights without sacrificing predictive power.
{"title":"Formal concept views for explainable boosting: A lattice-theoretic framework for Extreme Gradient Boosting and Gradient Boosting Models","authors":"Sherif Eneye Shuaib , Pakwan Riyapan , Jirapond Muangprathub","doi":"10.1016/j.iswa.2025.200569","DOIUrl":"10.1016/j.iswa.2025.200569","url":null,"abstract":"<div><div>Tree-based ensemble methods, such as Extreme Gradient Boosting (XGBoost) and Gradient Boosting models (GBM), are widely used for supervised learning due to their strong predictive capabilities. However, their complex architectures often hinder interpretability. This paper extends a lattice-theoretic framework originally developed for Random Forests to boosting algorithms, enabling a structured analysis of their internal logic via formal concept analysis (FCA).</div><div>We formally adapt four conceptual views: leaf, tree, tree predicate, and interordinal predicate to account for the sequential learning and optimization processes unique to boosting. Using the binary-class version of the car evaluation dataset from the OpenML CC18 benchmark suite, we conduct a systematic parameter study to examine how hyperparameters, such as tree depth and the number of trees, affect both model performance and conceptual complexity. Random Forest results from prior literature are used as a comparative baseline.</div><div>The results show that XGBoost yields the highest test accuracy, while GBM demonstrates greater stability in generalization error. Conceptually, boosting methods generate more compact and interpretable leaf views but preserve rich structural information in higher-level views. In contrast, Random Forests tend to produce denser and more redundant concept lattices. These trade-offs highlight how boosting methods, when interpreted through FCA, can strike a balance between performance and transparency.</div><div>Overall, this work contributes to explainable AI by demonstrating how lattice-based conceptual views can be systematically extended to complex boosting models, offering interpretable insights without sacrificing predictive power.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"27 ","pages":"Article 200569"},"PeriodicalIF":4.3,"publicationDate":"2025-08-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144907137","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}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号