Peristaltic pumps (PP), widely acknowledged for their benefits in pharmaceutical contexts, face challenges in achieving optimal dosing accuracy. This investigation contributes novel insights for the improvement of dosing precision, identifying how to apply AI models, specifically focusing on Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) neural networks over a realistic span of target volumes. To provide a more accurate representation of real-world performance, we consider a modified root mean square error metric () that directly compares dispensed volumes to target volumes. Based on this the study delves into two main methodologies: an iterative retraining method, called Online Training, and Pre-trained approach. Online Training shows best results, especially for volumes below 1.0 ml, achieving 38.4% improvement in and 31.6% in standard deviation (). Pre-trained models are faster and exhibit promising outcomes especially for volumes above 1.0 ml, with a three-features approach delivering the best performance (13.8% and 4.6% improvements in and , respectively). Overall, the findings highlight the effectiveness of iterative learning techniques, particularly for smaller dosage amounts, which complements the good performance of non-AI approaches for larger ones.
{"title":"Application of LSTM and GRU neural networks to improve peristaltic pump dosing accuracy","authors":"Davide Privitera , Stefano Bellissima , Sandro Bartolini","doi":"10.1016/j.iswa.2025.200571","DOIUrl":"10.1016/j.iswa.2025.200571","url":null,"abstract":"<div><div>Peristaltic pumps (PP), widely acknowledged for their benefits in pharmaceutical contexts, face challenges in achieving optimal dosing accuracy. This investigation contributes novel insights for the improvement of dosing precision, identifying how to apply AI models, specifically focusing on Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) neural networks over a realistic span of target volumes. To provide a more accurate representation of real-world performance, we consider a modified root mean square error metric (<span><math><mrow><mi>R</mi><mi>M</mi><mi>S</mi><msub><mrow><mi>E</mi></mrow><mrow><mi>P</mi><mi>P</mi></mrow></msub></mrow></math></span>) that directly compares dispensed volumes to target volumes. Based on this the study delves into two main methodologies: an iterative retraining method, called Online Training, and Pre-trained approach. Online Training shows best results, especially for volumes below 1.0 ml, achieving 38.4% improvement in <span><math><mrow><mi>R</mi><mi>M</mi><mi>S</mi><msub><mrow><mi>E</mi></mrow><mrow><mi>P</mi><mi>P</mi></mrow></msub></mrow></math></span> and 31.6% in standard deviation (<span><math><mrow><mi>S</mi><mi>T</mi><mi>D</mi></mrow></math></span>). Pre-trained models are faster and exhibit promising outcomes especially for volumes above 1.0 ml, with a three-features approach delivering the best performance (13.8% and 4.6% improvements in <span><math><mrow><mi>R</mi><mi>M</mi><mi>S</mi><msub><mrow><mi>E</mi></mrow><mrow><mi>P</mi><mi>P</mi></mrow></msub></mrow></math></span> and <span><math><mrow><mi>S</mi><mi>T</mi><mi>D</mi></mrow></math></span>, respectively). Overall, the findings highlight the effectiveness of iterative learning techniques, particularly for smaller dosage amounts, which complements the good performance of non-AI approaches for larger ones.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"27 ","pages":"Article 200571"},"PeriodicalIF":4.3,"publicationDate":"2025-08-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144895628","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 : 2025-08-19DOI: 10.1016/j.iswa.2025.200570
Salima Bourbia , Ayoub Karine , Aladine Chetouani , Mohammed El Hassouni , Maher Jridi
The increasing use of 3D point clouds in fields like virtual reality, robotics, and 3D gaming has made quality assessment a critical and essential task. Many no-reference point cloud quality assessment (NR-PCQA) methods fail to capture the critical relationship between geometric and color features, limiting their accuracy, and lacking their generalization capabilities. To address these challenges, we propose NSS-MDL, a NR-PCQA framework that integrates Natural Scene Statistics (NSS) into a multi-task deep learning architecture. The model is trained with two complementary tasks: the main task predicts the perceptual quality score, while the auxiliary task estimates NSS features. The main contribution of this work lies in the use of NSS estimation as an auxiliary task to enhance the capacity of deep learning-based models to represent both the naturalness and the degradation of point clouds, leading to more accurate and robust quality predictions Experimental evaluations on two large benchmark datasets, WPC and SJTU, demonstrate that NSS-MDL outperforms state-of-the-art methods in terms of correlation with subjective quality scores. The results highlight the robustness and generalizability of the proposed method across diverse datasets. The code of the NSS-MDL model will soon be publicly available on https://github.com/Salima-Bourbia/NSS-MDL.
{"title":"NSS-MDL: Natural Scene Statistics-guided multi-task deep learning for no-reference point cloud quality assessment","authors":"Salima Bourbia , Ayoub Karine , Aladine Chetouani , Mohammed El Hassouni , Maher Jridi","doi":"10.1016/j.iswa.2025.200570","DOIUrl":"10.1016/j.iswa.2025.200570","url":null,"abstract":"<div><div>The increasing use of 3D point clouds in fields like virtual reality, robotics, and 3D gaming has made quality assessment a critical and essential task. Many no-reference point cloud quality assessment (NR-PCQA) methods fail to capture the critical relationship between geometric and color features, limiting their accuracy, and lacking their generalization capabilities. To address these challenges, we propose NSS-MDL, a NR-PCQA framework that integrates Natural Scene Statistics (NSS) into a multi-task deep learning architecture. The model is trained with two complementary tasks: the main task predicts the perceptual quality score, while the auxiliary task estimates NSS features. The main contribution of this work lies in the use of NSS estimation as an auxiliary task to enhance the capacity of deep learning-based models to represent both the naturalness and the degradation of point clouds, leading to more accurate and robust quality predictions Experimental evaluations on two large benchmark datasets, WPC and SJTU, demonstrate that NSS-MDL outperforms state-of-the-art methods in terms of correlation with subjective quality scores. The results highlight the robustness and generalizability of the proposed method across diverse datasets. The code of the NSS-MDL model will soon be publicly available on <span><span>https://github.com/Salima-Bourbia/NSS-MDL</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"27 ","pages":"Article 200570"},"PeriodicalIF":4.3,"publicationDate":"2025-08-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144895745","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 : 2025-08-15DOI: 10.1016/j.iswa.2025.200572
Seyedmohammad Mousavian, Shah J Miah
Since studies of pattern recognition for detecting money laundering have overflowed with various outcomes, effective applications of artificial intelligence (AI) for delivering précised outcomes are still emerging. In this paper, we evaluate AI-based approaches for their performance measure (e.g., accuracy), data requirement, processing speed, and cost-effectiveness in detecting money laundering activities, find related gaps, and suggest possible courses of action. Adopting a smart literature review analysis, including PRISMA and a topic modeling technique, this study examines published peer-reviewed and conference articles from 2015 to June 2023. The study identifies dominant topics in the period, concluding that AI-based solutions have increasingly been deployed in detecting money laundering, though they face various challenges in application. It also emphasizes that AI solutions are required to be evaluated to measure their performance before applying to large-scale problem-solving.
{"title":"Review of artificial intelligence-based applications for money laundering detection","authors":"Seyedmohammad Mousavian, Shah J Miah","doi":"10.1016/j.iswa.2025.200572","DOIUrl":"10.1016/j.iswa.2025.200572","url":null,"abstract":"<div><div>Since studies of pattern recognition for detecting money laundering have overflowed with various outcomes, effective applications of artificial intelligence (AI) for delivering précised outcomes are still emerging. In this paper, we evaluate AI-based approaches for their performance measure (e.g., accuracy), data requirement, processing speed, and cost-effectiveness in detecting money laundering activities, find related gaps, and suggest possible courses of action. Adopting a smart literature review analysis, including PRISMA and a topic modeling technique, this study examines published peer-reviewed and conference articles from 2015 to June 2023. The study identifies dominant topics in the period, concluding that AI-based solutions have increasingly been deployed in detecting money laundering, though they face various challenges in application. It also emphasizes that AI solutions are required to be evaluated to measure their performance before applying to large-scale problem-solving.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"27 ","pages":"Article 200572"},"PeriodicalIF":4.3,"publicationDate":"2025-08-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144886316","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}
Fundamental analysis plays a critical role in equity investing, but its complexity has long limited the involvement of artificial intelligence (AI). Recent advances in large language models (LLMs), however, have opened new possibilities for AI to handle fundamental analysis. Despite this potential, leveraging LLMs to generate practically useful outputs remains a non-trivial challenge, and existing research is still in its early stages. This paper aims to enhance the performance of LLMs in fundamental analysis in a novel way, drawing inspiration from the practices of human analysts. We first propose a novel Autonomous Fundamental Analysis System (AutoFAS), which enables LLM agents to perform analyses on various topics of target companies. Next, we allow LLM agents to autonomously conduct research on specified companies with AutoFAS by exploring various topics they deem important, mimicking the experience accumulation of human analysts. Then, when presented with new research topics, the agents generate reports by referring to their accumulated analyses. Experiments show that, with AutoFAS, LLM agents can autonomously and logically explore various facets of target companies. The evaluation of their analysis on new research topics demonstrates that by drawing on accumulated analyses, they can naturally produce more unique and profound insights. This resembles the human process of generating novel ideas. Our work highlights a promising direction for applying LLMs in complex fundamental analysis, bridging the gap between human expertise and LLMs’ analysis.
{"title":"Can large language models autonomously generate unique and profound insights in fundamental analysis?","authors":"Tao Xu , Zhe Piao , Tadashi Mukai , Yuri Murayama , Kiyoshi Izumi","doi":"10.1016/j.iswa.2025.200566","DOIUrl":"10.1016/j.iswa.2025.200566","url":null,"abstract":"<div><div>Fundamental analysis plays a critical role in equity investing, but its complexity has long limited the involvement of artificial intelligence (AI). Recent advances in large language models (LLMs), however, have opened new possibilities for AI to handle fundamental analysis. Despite this potential, leveraging LLMs to generate practically useful outputs remains a non-trivial challenge, and existing research is still in its early stages. This paper aims to enhance the performance of LLMs in fundamental analysis in a novel way, drawing inspiration from the practices of human analysts. We first propose a novel Autonomous Fundamental Analysis System (AutoFAS), which enables LLM agents to perform analyses on various topics of target companies. Next, we allow LLM agents to autonomously conduct research on specified companies with AutoFAS by exploring various topics they deem important, mimicking the experience accumulation of human analysts. Then, when presented with new research topics, the agents generate reports by referring to their accumulated analyses. Experiments show that, with AutoFAS, LLM agents can autonomously and logically explore various facets of target companies. The evaluation of their analysis on new research topics demonstrates that by drawing on accumulated analyses, they can naturally produce more unique and profound insights. This resembles the human process of generating novel ideas. Our work highlights a promising direction for applying LLMs in complex fundamental analysis, bridging the gap between human expertise and LLMs’ analysis.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"27 ","pages":"Article 200566"},"PeriodicalIF":4.3,"publicationDate":"2025-08-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144861088","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 : 2025-08-11DOI: 10.1016/j.iswa.2025.200567
Haider A. Alwzwazy , Laith Alzubaidi , Zehui Zhao , Ahmed Saihood , Sabah Abdulazeez Jebur , Mohamed Manoufali , Omar Alnaseri , Jose Santamaria , Yuantong Gu
In recent years, deep learning-based multi-class and multi-task classification have gained significant attention across various domains of computer vision. However, current approaches often struggle to incorporate new classes efficiently due to the computational burden of retraining large neural networks from scratch. This limitation poses a significant obstacle to the deployment of deep learning models in real-world intelligent systems. Although continual learning has been proposed to overcome this challenge, it remains constrained by catastrophic forgetting. To address these limitations, this study introduces a new framework called Learning Without Retraining (LWR-Net), developed for multi-class and multi-task adaptation, allowing networks to adapt to new classes with minimal training requirements. Specifically, LWR-Net incorporates four key components: (i) task-guided self-supervised learning with a dual-attention mechanism to enhance feature generalisation and selection; (ii) task-based model fusion to improve feature representation and generalisation; (iii) multi-task learning to generalise classifiers across diverse tasks; and (iv) decision fusion of multiple classifiers to improve overall performance and reduce the likelihood of misclassification. LWR-Net was evaluated across diverse tasks to demonstrate its effectiveness in integrating new data, classes, or tasks. These include: (i) a medical case study detecting abnormalities in five distinct bone structures; (ii) a surveillance case study detecting violence in three different settings; and (iii) a geology case study identifying lateral changes in soil compaction using ground-penetrating radar across two datasets. The results show that LWR-Net achieves state-of-the-art performance across all three scenarios, successfully accommodates new learning objectives while preserving performance, eliminating the need for complete retraining cycles. Moreover, the use of gradient-weighted class activation mapping (Grad-CAM) confirmed that the models focused on relevant regions of interest. LWR-Net offers several benefits, including improved generalisation, enhanced performance, and the capacity to train on new data without catastrophic failures. The source code is publicly available at: https://github.com/LaithAlzubaidi/Learning-to-Adapt.
{"title":"LWR-Net: Learning without retraining for scalable multi-task adaptation and domain-agnostic generalisation","authors":"Haider A. Alwzwazy , Laith Alzubaidi , Zehui Zhao , Ahmed Saihood , Sabah Abdulazeez Jebur , Mohamed Manoufali , Omar Alnaseri , Jose Santamaria , Yuantong Gu","doi":"10.1016/j.iswa.2025.200567","DOIUrl":"10.1016/j.iswa.2025.200567","url":null,"abstract":"<div><div>In recent years, deep learning-based multi-class and multi-task classification have gained significant attention across various domains of computer vision. However, current approaches often struggle to incorporate new classes efficiently due to the computational burden of retraining large neural networks from scratch. This limitation poses a significant obstacle to the deployment of deep learning models in real-world intelligent systems. Although continual learning has been proposed to overcome this challenge, it remains constrained by catastrophic forgetting. To address these limitations, this study introduces a new framework called Learning Without Retraining (LWR-Net), developed for multi-class and multi-task adaptation, allowing networks to adapt to new classes with minimal training requirements. Specifically, LWR-Net incorporates four key components: (i) task-guided self-supervised learning with a dual-attention mechanism to enhance feature generalisation and selection; (ii) task-based model fusion to improve feature representation and generalisation; (iii) multi-task learning to generalise classifiers across diverse tasks; and (iv) decision fusion of multiple classifiers to improve overall performance and reduce the likelihood of misclassification. LWR-Net was evaluated across diverse tasks to demonstrate its effectiveness in integrating new data, classes, or tasks. These include: (i) a medical case study detecting abnormalities in five distinct bone structures; (ii) a surveillance case study detecting violence in three different settings; and (iii) a geology case study identifying lateral changes in soil compaction using ground-penetrating radar across two datasets. The results show that LWR-Net achieves state-of-the-art performance across all three scenarios, successfully accommodates new learning objectives while preserving performance, eliminating the need for complete retraining cycles. Moreover, the use of gradient-weighted class activation mapping (Grad-CAM) confirmed that the models focused on relevant regions of interest. LWR-Net offers several benefits, including improved generalisation, enhanced performance, and the capacity to train on new data without catastrophic failures. The source code is publicly available at: <span><span>https://github.com/LaithAlzubaidi/Learning-to-Adapt</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"27 ","pages":"Article 200567"},"PeriodicalIF":4.3,"publicationDate":"2025-08-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144828697","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}
The increasing complexity of educational recommendation systems, driven by the need to balance content relevance, learning outcomes, and diversity, demands advanced optimization solutions that overcome the limitations of traditional methods. As educational technology is exponentially improving, multi-objective optimization plays a vital role in adapting learning experiences to individual requirements. This study tackles the Gradient-Enhanced Evolutionary Multi-objective Optimization (GEEMOO) algorithm, which is considered as a hybrid framework that deals with three conflicting objectives: Relevance, Learning Outcomes, and Diversity. GEEMOO associates gradient-based methods for rapid integration with the correlative power of evolutionary strategies to deliver high-quality Pareto-optimal solutions. Extensive experimentation, using real-world datasets, has shown that GEEMOO consistently exceeded benchmark algorithms performance (NSGA-II and MOPSO) across key metrics, achieving greater Hypervolume, Generational Distance, and diversity indicators. While maintaining robust solution diversity, GEEMOO stands as an ideal solution for large-scale educational recommendation systems efficiency, requiring fewer fitness evaluations. GEEMOO showed better performance than NSGA-II and MOPSO in both convergence (Hypervolume: 0.85, Generational Distance: 0.02) and diversity (Spread Indicator: 0.88, Crowding Distance: 0.92). Although it required a bit more runtime (150 seconds compared to 120 seconds for NSGA-II), GEEMOO achieved this with fewer fitness evaluations (50,000 versus 60,000 for NSGA-II), highlighting its computational efficiency. The algorithm successfully balanced conflicting objectives, providing Pareto-optimal solutions that cater to various educational goals. This work traits GEEMOO’s adaptability and credibility to demonstrate how personalized learning models are adjusted, offering a solid groundwork for improving educational technology in both research and practice.
{"title":"Gradient-enhanced evolutionary multi-objective optimization (GEEMOO): Balancing relevance, learning outcomes, and diversity in educational recommendation systems","authors":"Youssef Jdidou , Souhaib Aammou , Hicham Er-radi , Ilias Aarab","doi":"10.1016/j.iswa.2025.200568","DOIUrl":"10.1016/j.iswa.2025.200568","url":null,"abstract":"<div><div>The increasing complexity of educational recommendation systems, driven by the need to balance content relevance, learning outcomes, and diversity, demands advanced optimization solutions that overcome the limitations of traditional methods. As educational technology is exponentially improving, multi-objective optimization plays a vital role in adapting learning experiences to individual requirements. This study tackles the Gradient-Enhanced Evolutionary Multi-objective Optimization (GEEMOO) algorithm, which is considered as a hybrid framework that deals with three conflicting objectives: Relevance, Learning Outcomes, and Diversity. GEEMOO associates gradient-based methods for rapid integration with the correlative power of evolutionary strategies to deliver high-quality Pareto-optimal solutions. Extensive experimentation, using real-world datasets, has shown that GEEMOO consistently exceeded benchmark algorithms performance (NSGA-II and MOPSO) across key metrics, achieving greater Hypervolume, Generational Distance, and diversity indicators. While maintaining robust solution diversity, GEEMOO stands as an ideal solution for large-scale educational recommendation systems efficiency, requiring fewer fitness evaluations. GEEMOO showed better performance than NSGA-II and MOPSO in both convergence (Hypervolume: 0.85, Generational Distance: 0.02) and diversity (Spread Indicator: 0.88, Crowding Distance: 0.92). Although it required a bit more runtime (150 seconds compared to 120 seconds for NSGA-II), GEEMOO achieved this with fewer fitness evaluations (50,000 versus 60,000 for NSGA-II), highlighting its computational efficiency. The algorithm successfully balanced conflicting objectives, providing Pareto-optimal solutions that cater to various educational goals. This work traits GEEMOO’s adaptability and credibility to demonstrate how personalized learning models are adjusted, offering a solid groundwork for improving educational technology in both research and practice.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"27 ","pages":"Article 200568"},"PeriodicalIF":4.3,"publicationDate":"2025-08-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144828696","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 : 2025-08-04DOI: 10.1016/j.iswa.2025.200564
Temitope Olubanjo Kehinde , Oluyinka J. Adedokun , Morenikeji Kabirat Kareem , Joseph Akpan , Oludolapo A. Olanrewaju
Accurate forecasting of high-volatility stock markets is critical for investors and policymakers, yet existing models struggle with computational inefficiency and noise sensitivity. This study introduces STL-ELM, a novel hybrid model combining Seasonal-Trend decomposition using LOESS (STL) and Extreme Learning Machine (ELM), to deliver unparalleled accuracy and speed. By decomposing stock data into trend, seasonal, and residual components, STL-ELM isolates multiscale features, while ELM’s lightweight architecture ensures rapid training and robust generalization, outperforming advanced techniques such as LSTM, GRU, and transformer variants in both prediction and trading simulations. With faster runtimes and minimal memory usage, STL-ELM is tailored for real-time trading applications and high-frequency financial forecasting, offering institutional investors, traders, and financial analysts a competitive edge in volatile markets. The hybrid nature of STL-ELM, which combines STL’s multiscale decomposition with ELM’s rapid learning, enhances its adaptability to various financial domains, including stocks, commodities, foreign exchange, and cryptocurrencies, by efficiently capturing domain-specific volatility patterns. This work not only sets a new standard for predictive accuracy in stock market modelling but also presents an invaluable tool for those navigating the complexities of modern financial markets.
{"title":"STL-ELM: A computationally efficient hybrid approach for predicting high volatility stock market","authors":"Temitope Olubanjo Kehinde , Oluyinka J. Adedokun , Morenikeji Kabirat Kareem , Joseph Akpan , Oludolapo A. Olanrewaju","doi":"10.1016/j.iswa.2025.200564","DOIUrl":"10.1016/j.iswa.2025.200564","url":null,"abstract":"<div><div>Accurate forecasting of high-volatility stock markets is critical for investors and policymakers, yet existing models struggle with computational inefficiency and noise sensitivity. This study introduces STL-ELM, a novel hybrid model combining Seasonal-Trend decomposition using LOESS (STL) and Extreme Learning Machine (ELM), to deliver unparalleled accuracy and speed. By decomposing stock data into trend, seasonal, and residual components, STL-ELM isolates multiscale features, while ELM’s lightweight architecture ensures rapid training and robust generalization, outperforming advanced techniques such as LSTM, GRU, and transformer variants in both prediction and trading simulations. With faster runtimes and minimal memory usage, STL-ELM is tailored for real-time trading applications and high-frequency financial forecasting, offering institutional investors, traders, and financial analysts a competitive edge in volatile markets. The hybrid nature of STL-ELM, which combines STL’s multiscale decomposition with ELM’s rapid learning, enhances its adaptability to various financial domains, including stocks, commodities, foreign exchange, and cryptocurrencies, by efficiently capturing domain-specific volatility patterns. This work not only sets a new standard for predictive accuracy in stock market modelling but also presents an invaluable tool for those navigating the complexities of modern financial markets.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"27 ","pages":"Article 200564"},"PeriodicalIF":4.3,"publicationDate":"2025-08-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144771355","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 classification of durian ripeness is essential for quality control and minimizing post-harvest losses. Manual inspection remains subjective and inconsistent, prompting the need for automated methods. We present a multi-modal approach that integrates Convolutional Neural Networks (CNNs) for image-based classification and Recurrent Neural Networks (RNNs) for automatic textual descriptions. Trained on 16,000 annotated images across four ripeness stages, the model achieved high classification accuracy (MobileNetV2: 95.50%) and superior captioning performance (ResNet101 + Bi-GRU: BLEU 0.9974, METEOR 0.9949, ROUGE 0.9164). While weighted summation fusion demonstrated superior performance, concatenation was ultimately chosen for its simplicity and real-world deployment feasibility. Statistical validation using one-way ANOVA () confirmed the significance of the findings. These results highlight the potential of the proposed multi-modal approach as a practical and interpretable framework for automated durian ripeness assessment.
{"title":"Multi-modal expert system for automated durian ripeness classification using deep learning","authors":"Santi Sukkasem, Watchareewan Jitsakul, Phayung Meesad","doi":"10.1016/j.iswa.2025.200563","DOIUrl":"10.1016/j.iswa.2025.200563","url":null,"abstract":"<div><div>Accurate classification of durian ripeness is essential for quality control and minimizing post-harvest losses. Manual inspection remains subjective and inconsistent, prompting the need for automated methods. We present a multi-modal approach that integrates Convolutional Neural Networks (CNNs) for image-based classification and Recurrent Neural Networks (RNNs) for automatic textual descriptions. Trained on 16,000 annotated images across four ripeness stages, the model achieved high classification accuracy (MobileNetV2: 95.50%) and superior captioning performance (ResNet101 + Bi-GRU: BLEU 0.9974, METEOR 0.9949, ROUGE 0.9164). While weighted summation fusion demonstrated superior performance, concatenation was ultimately chosen for its simplicity and real-world deployment feasibility. Statistical validation using one-way ANOVA (<span><math><mrow><mi>p</mi><mo><</mo><mn>0</mn><mo>.</mo><mn>05</mn></mrow></math></span>) confirmed the significance of the findings. These results highlight the potential of the proposed multi-modal approach as a practical and interpretable framework for automated durian ripeness assessment.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"27 ","pages":"Article 200563"},"PeriodicalIF":4.3,"publicationDate":"2025-07-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144766986","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 : 2025-07-31DOI: 10.1016/j.iswa.2025.200562
Serena Crisci , Valentina De Simone , Andrea Diana , Ferdinando Zullo
The increasing availability of visual data in fields such as archaeology has highlighted the need for automated image analysis tools. Ancient rock engravings, such as those in the Neolithic Domus de Janas tombs of Sardinia, are crucial cultural artifacts. However, their study is hindered by environmental degradation and the limitations of traditional analysis methods. This paper introduces a novel approach that employs a preprocessing method to isolate glyphs from their backgrounds, reducing the impact of wear and distortions caused by environmental factors such as lighting. Convolutional neural networks are then used to enhance the classification of glyphs in the preprocessed archaeological images. The refined data are processed using AlexNet, GoogLeNet, and EfficientNet neural networks, each trained to classify glyphs into distinct categories and to detect their geometric features. This method offers a more efficient and accurate way to analyze and preserve these cultural artifacts.
考古学等领域的视觉数据越来越多，这凸显了对自动图像分析工具的需求。古代岩石雕刻，如撒丁岛新石器时代Domus de Janas墓葬中的那些，是至关重要的文化文物。然而，它们的研究受到环境退化和传统分析方法的限制。本文介绍了一种新的方法，该方法采用预处理方法将字形从其背景中分离出来，减少了由光照等环境因素引起的磨损和扭曲的影响。然后使用卷积神经网络来增强预处理考古图像中的字形分类。精细化的数据使用AlexNet、GoogLeNet和effentnet神经网络进行处理，每个神经网络都经过训练，可以将字形分类为不同的类别，并检测其几何特征。这种方法为分析和保存这些文物提供了一种更有效、更准确的方法。
{"title":"Neural network for archaeological glyph detection","authors":"Serena Crisci , Valentina De Simone , Andrea Diana , Ferdinando Zullo","doi":"10.1016/j.iswa.2025.200562","DOIUrl":"10.1016/j.iswa.2025.200562","url":null,"abstract":"<div><div>The increasing availability of visual data in fields such as archaeology has highlighted the need for automated image analysis tools. Ancient rock engravings, such as those in the Neolithic Domus de Janas tombs of Sardinia, are crucial cultural artifacts. However, their study is hindered by environmental degradation and the limitations of traditional analysis methods. This paper introduces a novel approach that employs a preprocessing method to isolate glyphs from their backgrounds, reducing the impact of wear and distortions caused by environmental factors such as lighting. Convolutional neural networks are then used to enhance the classification of glyphs in the preprocessed archaeological images. The refined data are processed using AlexNet, GoogLeNet, and EfficientNet neural networks, each trained to classify glyphs into distinct categories and to detect their geometric features. This method offers a more efficient and accurate way to analyze and preserve these cultural artifacts.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"27 ","pages":"Article 200562"},"PeriodicalIF":4.3,"publicationDate":"2025-07-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144766985","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 : 2025-07-30DOI: 10.1016/j.iswa.2025.200560
Wonjik Kim , Gaku Kutsuzawa , Michiyo Maruyama
Wearable devices enable the continuous acquisition of physiological signals, offering the potential for real-time emotion monitoring in daily life. However, emotion recognition remains challenging due to individual differences, label ambiguity, and limited annotated data. This study proposes a lightweight, cluster-guided attention model for binary emotion recognition (positive vs. negative) and forecasting (up to two hours ahead) from wearable signals such as heart rate and step count. To improve generalization, we leverage unsupervised clustering in the latent space and integrate cross-species pretraining using structured behavioral and physiological data from mice. Our framework reduces annotation burden through an emoji-based self-report interface and performs both within- and across-subject validation. Experimental results on human wearable data demonstrate that our method outperforms classical and lightweight deep learning baselines in both accuracy and macro-F1 score, achieving approximately 74.4% accuracy (macro-F1: 71.5%) for current emotion recognition, 72.9% accuracy (macro-F1: 70.7%) for 1-h forecasting, and 65.5% accuracy (macro-F1: 63.0%) for 2-h forecasting. Moreover, mouse-based pretraining yields consistent performance gains, especially at longer-horizon prediction tasks. These findings suggest that biologically informed attention mechanisms and cross-domain knowledge transfer can significantly enhance emotion modeling from low-resource wearable data.
{"title":"Emotion recognition and forecasting from wearable data via cluster-guided attention with cross-species pretraining","authors":"Wonjik Kim , Gaku Kutsuzawa , Michiyo Maruyama","doi":"10.1016/j.iswa.2025.200560","DOIUrl":"10.1016/j.iswa.2025.200560","url":null,"abstract":"<div><div>Wearable devices enable the continuous acquisition of physiological signals, offering the potential for real-time emotion monitoring in daily life. However, emotion recognition remains challenging due to individual differences, label ambiguity, and limited annotated data. This study proposes a lightweight, cluster-guided attention model for binary emotion recognition (positive vs. negative) and forecasting (up to two hours ahead) from wearable signals such as heart rate and step count. To improve generalization, we leverage unsupervised clustering in the latent space and integrate cross-species pretraining using structured behavioral and physiological data from mice. Our framework reduces annotation burden through an emoji-based self-report interface and performs both within- and across-subject validation. Experimental results on human wearable data demonstrate that our method outperforms classical and lightweight deep learning baselines in both accuracy and macro-F1 score, achieving approximately 74.4% accuracy (macro-F1: 71.5%) for current emotion recognition, 72.9% accuracy (macro-F1: 70.7%) for 1-h forecasting, and 65.5% accuracy (macro-F1: 63.0%) for 2-h forecasting. Moreover, mouse-based pretraining yields consistent performance gains, especially at longer-horizon prediction tasks. These findings suggest that biologically informed attention mechanisms and cross-domain knowledge transfer can significantly enhance emotion modeling from low-resource wearable data.</div></div>","PeriodicalId":100684,"journal":{"name":"Intelligent Systems with Applications","volume":"27 ","pages":"Article 200560"},"PeriodicalIF":4.3,"publicationDate":"2025-07-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144750473","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