Pub Date : 2026-01-10DOI: 10.1016/j.ins.2026.123105
Yadong Zhao, Jie Lian, Dong Wang
Cooperative navigation of multiple robots in complex and dynamic environments has been regarded as a key requirement for many real-world applications. Consequently, efficient and safe navigation in such environments is considered to rely heavily on effective information processing and decision-making. However, the effectiveness of existing navigation approaches is significantly restricted by environmental uncertainty and the challenges associated with processing heterogeneous information. To address these problems, an adaptive risk-averse reinforcement learning method, termed ARA-MDSAC, is proposed, in which an adaptive risk-averse strategy is incorporated together with a bidirectional Mamba module for feature extraction. In this strategy, the full return distribution is modeled using a quantile regression mechanism. The environmental uncertainty is estimated via a random network distillation module and mapped to a distortion function, enabling adaptive risk preference adjustment and risk-averse decision-making under environmental uncertainty. A feature extraction module based on bidirectional Mamba is designed to encode heterogeneous information into sequential representations, capture bidirectional contextual dependencies, and enable efficient feature fusion and extraction. Experiments are conducted to demonstrate the effectiveness of the proposed method in complex environments.
{"title":"Adaptive risk-averse reinforcement learning for cooperative navigation of multiple robots","authors":"Yadong Zhao, Jie Lian, Dong Wang","doi":"10.1016/j.ins.2026.123105","DOIUrl":"10.1016/j.ins.2026.123105","url":null,"abstract":"<div><div>Cooperative navigation of multiple robots in complex and dynamic environments has been regarded as a key requirement for many real-world applications. Consequently, efficient and safe navigation in such environments is considered to rely heavily on effective information processing and decision-making. However, the effectiveness of existing navigation approaches is significantly restricted by environmental uncertainty and the challenges associated with processing heterogeneous information. To address these problems, an adaptive risk-averse reinforcement learning method, termed ARA-MDSAC, is proposed, in which an adaptive risk-averse strategy is incorporated together with a bidirectional Mamba module for feature extraction. In this strategy, the full return distribution is modeled using a quantile regression mechanism. The environmental uncertainty is estimated via a random network distillation module and mapped to a distortion function, enabling adaptive risk preference adjustment and risk-averse decision-making under environmental uncertainty. A feature extraction module based on bidirectional Mamba is designed to encode heterogeneous information into sequential representations, capture bidirectional contextual dependencies, and enable efficient feature fusion and extraction. Experiments are conducted to demonstrate the effectiveness of the proposed method in complex environments.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"736 ","pages":"Article 123105"},"PeriodicalIF":6.8,"publicationDate":"2026-01-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146038884","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-09DOI: 10.1016/j.ins.2026.123101
Meiji Cui , Wenping Wang , Shuwei Zhu , Wei Fang , Mengchu Tian
Problem-transformation-based methods have been acknowledged as effective approaches for addressing Large-Scale Multi-objective Optimization Problems (LSMOPs), as they enable the conversion of the original high-dimensional search space into a relatively low-dimensional one. Nevertheless, such methods often exhibit an extremely rapid convergence rate, which is concomitantly accompanied by a significant loss of diversity. Within this context, the selection of reference solutions—upon which the problem transformation is constructed—plays a pivotal role in influencing population diversity. To address these limitations, this paper proposes a novel Reference Solution Selection strategy based on Auxiliary Spaces (RSAS) designed to preserve population diversity. Specifically, orthogonal bases are constructed utilizing extreme feasible solutions positioned on the coordinate axes, with each basis spanning a distinct subspace. The ensemble of subspaces generated by multiple such orthogonal bases is defined as auxiliary spaces. These auxiliary spaces, characterized by their relatively uniform distribution throughout the original decision space, facilitate the algorithm in generating a reference solution set with substantially enhanced diversity. Notably, RSAS can be seamlessly integrated into any problem-transformation-based framework that incorporates a reference selection mechanism, thereby improving their overall performance. To demonstrate the efficacy of RSAS, comprehensive experimental investigations are conducted on benchmark suites with dimensions ranging from 500 to 5000. The experimental results conclusively validate the superiority of RSAS in comparison to other state-of-the-art algorithms on most instances.
{"title":"A novel reference solution selection strategy based on auxiliary space for large-scale multi-objective optimization","authors":"Meiji Cui , Wenping Wang , Shuwei Zhu , Wei Fang , Mengchu Tian","doi":"10.1016/j.ins.2026.123101","DOIUrl":"10.1016/j.ins.2026.123101","url":null,"abstract":"<div><div>Problem-transformation-based methods have been acknowledged as effective approaches for addressing Large-Scale Multi-objective Optimization Problems (LSMOPs), as they enable the conversion of the original high-dimensional search space into a relatively low-dimensional one. Nevertheless, such methods often exhibit an extremely rapid convergence rate, which is concomitantly accompanied by a significant loss of diversity. Within this context, the selection of reference solutions—upon which the problem transformation is constructed—plays a pivotal role in influencing population diversity. To address these limitations, this paper proposes a novel Reference Solution Selection strategy based on Auxiliary Spaces (RSAS) designed to preserve population diversity. Specifically, orthogonal bases are constructed utilizing extreme feasible solutions positioned on the coordinate axes, with each basis spanning a distinct subspace. The ensemble of subspaces generated by multiple such orthogonal bases is defined as auxiliary spaces. These auxiliary spaces, characterized by their relatively uniform distribution throughout the original decision space, facilitate the algorithm in generating a reference solution set with substantially enhanced diversity. Notably, RSAS can be seamlessly integrated into any problem-transformation-based framework that incorporates a reference selection mechanism, thereby improving their overall performance. To demonstrate the efficacy of RSAS, comprehensive experimental investigations are conducted on benchmark suites with dimensions ranging from 500 to 5000. The experimental results conclusively validate the superiority of RSAS in comparison to other state-of-the-art algorithms on most instances.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"735 ","pages":"Article 123101"},"PeriodicalIF":6.8,"publicationDate":"2026-01-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145978799","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-09DOI: 10.1016/j.ins.2026.123092
Wenfu Huang , Jie Zhao , Ran Wei , Jingyuan Yang , Jing Ruan , Li Fan , Xinwen Zhou
Defects arising during photovoltaic (PV) cell manufacturing critically compromise performance and operational safety. Existing computer vision-based detection methods struggle with defects characterized by small scales, dense distributions, and background confusion. To address this, we propose MRFP-TDG, a Detection Transformer-based detector featuring an optimized hybrid encoder for enhanced fine-grained feature extraction and background suppression. Specifically, our Multi-scale Receptive Field Projection (MRFP) module leverages channel-split depthwise separable convolutions with multi-scale kernels to project backbone features into tokens while preserving spatial relationships, significantly improving small-defect detection. The Token-Driven Gathering (TDG) module further integrates spatial attention to fuse multi-scale tokens, compensating for tokenization-induced spatial information loss while suppressing background noise. Furthermore, a relational position embedding mechanism in the decoder models positional relationships of bounding boxes across layers, accelerating convergence during iterative refinement. Evaluated on the public PVEL-AD dataset, MRFP-TDG achieves 95.1% mAP@50 in nine-category defect detection, outperforming state-of-the-art PV defect detectors in both accuracy and efficiency. Specifically, it surpasses the baseline model by 1.0% mAP while requiring only 59.6% of the computational cost compared with the best-performing SOTA method.
{"title":"MRFP-TDG: A detection transformer with hybrid encoder and position-aware decoder for photovoltaic cell defect detection","authors":"Wenfu Huang , Jie Zhao , Ran Wei , Jingyuan Yang , Jing Ruan , Li Fan , Xinwen Zhou","doi":"10.1016/j.ins.2026.123092","DOIUrl":"10.1016/j.ins.2026.123092","url":null,"abstract":"<div><div>Defects arising during photovoltaic (PV) cell manufacturing critically compromise performance and operational safety. Existing computer vision-based detection methods struggle with defects characterized by small scales, dense distributions, and background confusion. To address this, we propose MRFP-TDG, a Detection Transformer-based detector featuring an optimized hybrid encoder for enhanced fine-grained feature extraction and background suppression. Specifically, our Multi-scale Receptive Field Projection (MRFP) module leverages channel-split depthwise separable convolutions with multi-scale kernels to project backbone features into tokens while preserving spatial relationships, significantly improving small-defect detection. The Token-Driven Gathering (TDG) module further integrates spatial attention to fuse multi-scale tokens, compensating for tokenization-induced spatial information loss while suppressing background noise. Furthermore, a relational position embedding mechanism in the decoder models positional relationships of bounding boxes across layers, accelerating convergence during iterative refinement. Evaluated on the public PVEL-AD dataset, MRFP-TDG achieves 95.1% mAP@50 in nine-category defect detection, outperforming state-of-the-art PV defect detectors in both accuracy and efficiency. Specifically, it surpasses the baseline model by 1.0% mAP while requiring only 59.6% of the computational cost compared with the best-performing SOTA method.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"735 ","pages":"Article 123092"},"PeriodicalIF":6.8,"publicationDate":"2026-01-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145978797","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-09DOI: 10.1016/j.ins.2026.123103
Hoang-Thang Ta , Duy-Quy Thai , Abu Bakar Siddiqur Rahman , Grigori Sidorov , Alexander Gelbukh
In this paper, we introduce FC-KAN, a Kolmogorov-Arnold Network (KAN) that leverages combinations of popular mathematical functions such as B-splines, wavelets, and radial basis functions on low-dimensional data through element-wise operations. We explore several methods for combining the outputs of these functions, including sum, element-wise product, the addition of sum and element-wise product, representations of quadratic and cubic functions, concatenation, linear transformation of the concatenated output, and others. In our experiments, we compare FC-KAN with a multi-layer perceptron network (MLP) and other existing KANs, such as BSRBF-KAN, EfficientKAN, FastKAN, and FasterKAN, on the MNIST and Fashion-MNIST datasets. Two variants of FC-KAN, which use a combination of outputs from B-splines and Difference of Gaussians (DoG) and from B-splines and linear transformations in the form of a quadratic function, outperformed all other models on the average of 5 independent training runs. However, FC-KAN still has limitations, including challenges with parameter scalability and efficiency, as well as limited capability compared to CNNs when handling multi-channel datasets such as CIFAR-10 and CIFAR-100. We expect that FC-KAN can leverage function combinations to design future KANs. Our repository is publicly available at: https://github.com/hoangthangta/FC_KAN.
{"title":"FC-KAN: Function combinations in Kolmogorov-Arnold networks","authors":"Hoang-Thang Ta , Duy-Quy Thai , Abu Bakar Siddiqur Rahman , Grigori Sidorov , Alexander Gelbukh","doi":"10.1016/j.ins.2026.123103","DOIUrl":"10.1016/j.ins.2026.123103","url":null,"abstract":"<div><div>In this paper, we introduce FC-KAN, a Kolmogorov-Arnold Network (KAN) that leverages combinations of popular mathematical functions such as B-splines, wavelets, and radial basis functions on low-dimensional data through element-wise operations. We explore several methods for combining the outputs of these functions, including sum, element-wise product, the addition of sum and element-wise product, representations of quadratic and cubic functions, concatenation, linear transformation of the concatenated output, and others. In our experiments, we compare FC-KAN with a multi-layer perceptron network (MLP) and other existing KANs, such as BSRBF-KAN, EfficientKAN, FastKAN, and FasterKAN, on the MNIST and Fashion-MNIST datasets. Two variants of FC-KAN, which use a combination of outputs from B-splines and Difference of Gaussians (DoG) and from B-splines and linear transformations in the form of a quadratic function, outperformed all other models on the average of 5 independent training runs. However, FC-KAN still has limitations, including challenges with parameter scalability and efficiency, as well as limited capability compared to CNNs when handling multi-channel datasets such as CIFAR-10 and CIFAR-100. We expect that FC-KAN can leverage function combinations to design future KANs. Our repository is publicly available at: <span><span>https://github.com/hoangthangta/FC_KAN</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"736 ","pages":"Article 123103"},"PeriodicalIF":6.8,"publicationDate":"2026-01-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145981580","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-08DOI: 10.1016/j.ins.2026.123100
Congxin Wei , Zidong Quan , Yaxin Su , Haikuo Pang , Lei Wang , Shuhaida Mohamed Shuhidan , Safwan Mahmood Al-Selwi , Mohd Fadzil Hassan
GNSS time series are characterized by nonstationarity, high noise, and multi-scale signal mixing, posing challenges for accurate modeling with traditional approaches. This study proposes a hybrid prediction framework that integrates CEEMDAN, frequency-domain clustering reconstruction, and a secondary variational mode decomposition (VMD). The proposed model, CEDV-TCN-GTAF, employs a temporal convolutional network (TCN) for shared feature extraction and uses gated recurrent units (GRU) and Transformer networks in parallel to capture temporal dependencies and long-range correlations. An adaptive weighting mechanism dynamically fuses their outputs. In the decomposition stage, spectral features extracted via FFT are used as the input to K-means to cluster intrinsic mode functions (IMFs) into high-, mid-, and low-frequency groups. High-frequency components are further decomposed using DE-optimized VMD to ensure structural uniformity. Experiments on two regional GNSS datasets, YNYL and SCJU, validate the method’s effectiveness. On YNYL, compared with GRU, R2 increased by 71.53%, SMAPE decreased by 69.36%, and MAE and RMSE were reduced by 33.02% and 32.23%. Relative to the strongest baseline, CEDV-GTAF, the method further achieves notable gains, with MAE and MSE improved by 30.64% and 50.82%, and SMAPE reduced by 22.16%. On SCJU, compared with CEEMDAN-TCN-GTAF, R2 increased by 16.61%, SMAPE decreased by 39.07%, and MAE and RMSE were reduced by 52.51% and 55.38%. When benchmarked against the strongest baseline, CEDV-GTAF, MAE, MSE, and SMAPE are additionally improved by 10.18%, 18.84%, and 9.23%. Further, against CEDV-TCN-GT-AVG, MAE and RMSE decreased by 23.57% and 42.34%. These results demonstrate the proposed model’s robustness and superiority in forecasting nonstationary GNSS sequences and highlight its potential for geohazard applications such as earthquake monitoring and landslide early warning.
{"title":"A frequency-domain decomposition and TCN-GTAF fusion framework for GNSS sequence forecasting","authors":"Congxin Wei , Zidong Quan , Yaxin Su , Haikuo Pang , Lei Wang , Shuhaida Mohamed Shuhidan , Safwan Mahmood Al-Selwi , Mohd Fadzil Hassan","doi":"10.1016/j.ins.2026.123100","DOIUrl":"10.1016/j.ins.2026.123100","url":null,"abstract":"<div><div>GNSS time series are characterized by nonstationarity, high noise, and multi-scale signal mixing, posing challenges for accurate modeling with traditional approaches. This study proposes a hybrid prediction framework that integrates CEEMDAN, frequency-domain clustering reconstruction, and a secondary variational mode decomposition (VMD). The proposed model, CEDV-TCN-GTAF, employs a temporal convolutional network (TCN) for shared feature extraction and uses gated recurrent units (GRU) and Transformer networks in parallel to capture temporal dependencies and long-range correlations. An adaptive weighting mechanism dynamically fuses their outputs. In the decomposition stage, spectral features extracted via FFT are used as the input to K-means to cluster intrinsic mode functions (IMFs) into high-, mid-, and low-frequency groups. High-frequency components are further decomposed using DE-optimized VMD to ensure structural uniformity. Experiments on two regional GNSS datasets, YNYL and SCJU, validate the method’s effectiveness. On YNYL, compared with GRU, R<sup>2</sup> increased by 71.53%, SMAPE decreased by 69.36%, and MAE and RMSE were reduced by 33.02% and 32.23%. Relative to the strongest baseline, CEDV-GTAF, the method further achieves notable gains, with MAE and MSE improved by 30.64% and 50.82%, and SMAPE reduced by 22.16%. On SCJU, compared with CEEMDAN-TCN-GTAF, R<sup>2</sup> increased by 16.61%, SMAPE decreased by 39.07%, and MAE and RMSE were reduced by 52.51% and 55.38%. When benchmarked against the strongest baseline, CEDV-GTAF, MAE, MSE, and SMAPE are additionally improved by 10.18%, 18.84%, and 9.23%. Further, against CEDV-TCN-GT-AVG, MAE and RMSE decreased by 23.57% and 42.34%. These results demonstrate the proposed model’s robustness and superiority in forecasting nonstationary GNSS sequences and highlight its potential for geohazard applications such as earthquake monitoring and landslide early warning.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"735 ","pages":"Article 123100"},"PeriodicalIF":6.8,"publicationDate":"2026-01-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145978800","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-08DOI: 10.1016/j.ins.2026.123097
Ze Li , Rui Wang , Dianbo Ruan , Zhijun Qiao , Bin Huang
Accurate capacity estimation is crucial for the safety of supercapacitors. However, the effectiveness of the small-sample-data-driven capacity estimation is limited owing to the difficulty in obtaining sufficient capacity labels under complex application conditions. To improve the capacity estimation of supercapacitors under various conditions, an adaptive intelligent transfer learning method is first presented for the state estimation of supercapacitors. First, the aging and capacity degradation mechanisms were analyzed to extract the charge time, discharge time, direct current internal resistance, and median discharge voltage as features of the data-driven model. Based on the acquired features, the domain adaptation transfer learning framework with the margin disparity discrepancy was introduced, which improves knowledge transferring from the source domain to the target domain. Furthermore, considering the noise of feature data under different working conditions, an adaptive Gaussian filtering noise reduction-based convolutional neural network was integrated to enhance the feature-based knowledge transfer for different working conditions. The effectiveness of the proposed method was verified via an AVX R-Type Lug Terminal Supercapacitor. The proposed method can obtain high accuracy with adaptive noise reduction and a transfer learning domain adaptation combination strategy with mean absolute relative error below 0.15 and a mean absolute error and root mean square error of approximately 0.10, which demonstrates the effectiveness of the dual-driven strategy integrating denoising and transfer learning for domain adaptation.
{"title":"Transfer learning-based capacity estimation of supercapacitors adapted to different working conditions","authors":"Ze Li , Rui Wang , Dianbo Ruan , Zhijun Qiao , Bin Huang","doi":"10.1016/j.ins.2026.123097","DOIUrl":"10.1016/j.ins.2026.123097","url":null,"abstract":"<div><div>Accurate capacity estimation is crucial for the safety of supercapacitors. However, the effectiveness of the small-sample-data-driven capacity estimation is limited owing to the difficulty in obtaining sufficient capacity labels under complex application conditions. To improve the capacity estimation of supercapacitors under various conditions, an adaptive intelligent transfer learning method is first presented for the state estimation of supercapacitors. First, the aging and capacity degradation mechanisms were analyzed to extract the charge time, discharge time, direct current internal resistance, and median discharge voltage as features of the data-driven model. Based on the acquired features, the domain adaptation transfer learning framework with the margin disparity discrepancy was introduced, which improves knowledge transferring from the source domain to the target domain. Furthermore, considering the noise of feature data under different working conditions, an adaptive Gaussian filtering noise reduction-based convolutional neural network was integrated to enhance the feature-based knowledge transfer for different working conditions. The effectiveness of the proposed method was verified via an AVX R-Type Lug Terminal Supercapacitor. The proposed method can obtain high accuracy with adaptive noise reduction and a transfer learning domain adaptation combination strategy with mean absolute relative error below 0.15 and a mean absolute error and root mean square error of approximately 0.10, which demonstrates the effectiveness of the dual-driven strategy integrating denoising and transfer learning for domain adaptation.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"735 ","pages":"Article 123097"},"PeriodicalIF":6.8,"publicationDate":"2026-01-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145978689","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-08DOI: 10.1016/j.ins.2026.123094
Kwangeun Cho, Jungmin Lee, Seoung Bum Kim
Anomaly detection in irregular multivariate time series (IMTS) plays a crucial role in diverse applications such as fault prediction in industrial systems, energy management, and medical diagnosis. However, most existing methods are developed under the assumption of regularly sampled data, which makes them insufficient for real-world scenarios characterized by irregular sampling and missing values. We aim to develop an unsupervised framework specifically tailored for anomaly detection in IMTS. We propose SFAFormer, a framework that adopts a sampling frequency-aware (SFA) embedding to convert irregular time series into fixed-length vectors and a dual transformer encoder architecture to jointly capture temporal dependencies and inter-variable interactions. Furthermore, the input sequence is divided into patches, and both inter-patch and intra-patch relationships are modeled to effectively identify anomaly patterns. Experimental evaluations on four benchmark datasets (PSM, SMD, SWAT and GECCO) show that SFAFormer consistently outperforms existing approaches, achieving F1-score improvements of up to 30%p and AUROC gains of up to 18.9%p while maintaining robustness under diverse irregular sampling conditions. These findings demonstrate that SFAFormer provides an effective and practical solution for anomaly detection in IMTS.
{"title":"SFAFormer: Sampling Frequency-Aware Transformer Specialized for Unsupervised Anomaly Detection in Irregular Multivariate Time Series","authors":"Kwangeun Cho, Jungmin Lee, Seoung Bum Kim","doi":"10.1016/j.ins.2026.123094","DOIUrl":"10.1016/j.ins.2026.123094","url":null,"abstract":"<div><div>Anomaly detection in irregular multivariate time series (IMTS) plays a crucial role in diverse applications such as fault prediction in industrial systems, energy management, and medical diagnosis. However, most existing methods are developed under the assumption of regularly sampled data, which makes them insufficient for real-world scenarios characterized by irregular sampling and missing values. We aim to develop an unsupervised framework specifically tailored for anomaly detection in IMTS. We propose SFAFormer, a framework that adopts a sampling frequency-aware (SFA) embedding to convert irregular time series into fixed-length vectors and a dual transformer encoder architecture to jointly capture temporal dependencies and inter-variable interactions. Furthermore, the input sequence is divided into patches, and both inter-patch and intra-patch relationships are modeled to effectively identify anomaly patterns. Experimental evaluations on four benchmark datasets (PSM, SMD, SWAT and GECCO) show that SFAFormer consistently outperforms existing approaches, achieving F1-score improvements of up to 30%p and AUROC gains of up to 18.9%p while maintaining robustness under diverse irregular sampling conditions. These findings demonstrate that SFAFormer provides an effective and practical solution for anomaly detection in IMTS.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"736 ","pages":"Article 123094"},"PeriodicalIF":6.8,"publicationDate":"2026-01-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146038888","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-07DOI: 10.1016/j.ins.2026.123096
Wei Li, Bin Pang
As two non-associative binary fuzzy logical operators, overlap and grouping functions have been integrated into rough set theory. However, research on covering-based variable precision fuzzy rough sets (CVPFRSs) from the perspective of overlap and grouping functions remains limited, leaving several gaps to be addressed. To fill this gap, we propose novel CVPFRS models based on overlap and grouping functions, referred to as -CVPFRSs, and develop corresponding multi-attribute decision-making (MADM) methods. First, by employing residual implications and coimplications derived from overlap and grouping functions, we construct four distinct -CVPFRS models and systematically investigate their theoretical properties, with particular emphasis on their comparability property. Subsequently, building upon the traditional TOPSIS method, we propose two MADM methods grounded in the -CVPFRS models. Finally, we validate the proposed methods through a numerical case study. Comparative analyses with benchmark approaches demonstrate the validity, reliability, and practical effectiveness of the proposed methods in material selection for bone grafting.
{"title":"Covering-based (O,G)-variable precision fuzzy rough set and its application in decision-making","authors":"Wei Li, Bin Pang","doi":"10.1016/j.ins.2026.123096","DOIUrl":"10.1016/j.ins.2026.123096","url":null,"abstract":"<div><div>As two non-associative binary fuzzy logical operators, overlap and grouping functions have been integrated into rough set theory. However, research on covering-based variable precision fuzzy rough sets (CVPFRSs) from the perspective of overlap and grouping functions remains limited, leaving several gaps to be addressed. To fill this gap, we propose novel CVPFRS models based on overlap and grouping functions, referred to as <span><math><mo>(</mo><mi>O</mi><mo>,</mo><mi>G</mi><mo>)</mo></math></span>-CVPFRSs, and develop corresponding multi-attribute decision-making (MADM) methods. First, by employing residual implications and coimplications derived from overlap and grouping functions, we construct four distinct <span><math><mo>(</mo><mi>O</mi><mo>,</mo><mi>G</mi><mo>)</mo></math></span>-CVPFRS models and systematically investigate their theoretical properties, with particular emphasis on their comparability property. Subsequently, building upon the traditional TOPSIS method, we propose two MADM methods grounded in the <span><math><mo>(</mo><mi>O</mi><mo>,</mo><mi>G</mi><mo>)</mo></math></span>-CVPFRS models. Finally, we validate the proposed methods through a numerical case study. Comparative analyses with benchmark approaches demonstrate the validity, reliability, and practical effectiveness of the proposed methods in material selection for bone grafting.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"735 ","pages":"Article 123096"},"PeriodicalIF":6.8,"publicationDate":"2026-01-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145978686","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-07DOI: 10.1016/j.ins.2026.123080
Huanyu Wan , Shouzhen Zeng
Given that current research on multi-criteria group decision-making (MCGDM) methods primarily focuses on integrating deep learning algorithms, such as graph neural networks, there has been relatively little exploration of unsupervised learning algorithms. This research investigates this matter by introducing the concept of the Gaussian mixture model (GMM) in the context of group decision-making. As an unsupervised learning approach, GMM effectively captures the latent relationships among alternatives through cluster analysis and optimizes decision-making information using the expectation-maximization (EM) algorithm, thereby enhancing the construction of decision matrices and the ranking of alternatives. First, the EM algorithm within GMM is employed to process the raw information in group decision-making, resulting in an updated decision matrix. Next, to address uncertainties and fuzzy information in decision-making, the MCGDM problem is mapped into a Pythagorean fuzzy environment, where an innovative entropy measure is proposed to calculate the criteria weights using the entropy weight method. Additionally, a novel distance measure is developed and incorporated into the grey relational analysis (GRA)-TOPSIS approach, enabling comprehensive evaluation by reflecting the mutual relationships among alternatives and their closeness to the ideal option. Finally, an empirical evaluation of “zero-waste city” development demonstrates the practical applicability and effectiveness of the proposed approach.
{"title":"Gaussian mixture model-based Pythagorean fuzzy multi-criteria group decision-making method and its application in “zero-waste city” evaluation","authors":"Huanyu Wan , Shouzhen Zeng","doi":"10.1016/j.ins.2026.123080","DOIUrl":"10.1016/j.ins.2026.123080","url":null,"abstract":"<div><div>Given that current research on multi-criteria group decision-making (MCGDM) methods primarily focuses on integrating deep learning algorithms, such as graph neural networks, there has been relatively little exploration of unsupervised learning algorithms. This research investigates this matter by introducing the concept of the Gaussian mixture model (GMM) in the context of group decision-making. As an unsupervised learning approach, GMM effectively captures the latent relationships among alternatives through cluster analysis and optimizes decision-making information using the expectation-maximization (EM) algorithm, thereby enhancing the construction of decision matrices and the ranking of alternatives. First, the EM algorithm within GMM is employed to process the raw information in group decision-making, resulting in an updated decision matrix. Next, to address uncertainties and fuzzy information in decision-making, the MCGDM problem is mapped into a Pythagorean fuzzy environment, where an innovative entropy measure is proposed to calculate the criteria weights using the entropy weight method. Additionally, a novel distance measure is developed and incorporated into the grey relational analysis (GRA)-TOPSIS approach, enabling comprehensive evaluation by reflecting the mutual relationships among alternatives and their closeness to the ideal option. Finally, an empirical evaluation of “zero-waste city” development demonstrates the practical applicability and effectiveness of the proposed approach.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"735 ","pages":"Article 123080"},"PeriodicalIF":6.8,"publicationDate":"2026-01-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145978794","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-07DOI: 10.1016/j.ins.2025.123046
Peide Liu , Abbas Ali , Noor Rehman , Areej Qadeer
The three-way decision (3WD) model has gained widespread application in multi-attribute decision-making. However, existing models often neglect the variability in decision-makers minimum acceptance levels and risk attitudes across different criteria. With growing complexity and uncertainty in decision contexts, accurately capturing evaluation values remains a key challenge. In this paper, we address uncertainty in multi-attribute decision-making by introducing a novel type of fuzzy preference relation and developing collective fuzzy preference relations derived from multisource fuzzy information. To reduce the influence of subjective factors, we propose a new approach for calculating conditional probabilities based on collective fuzzy preferences and attribute weights. Furthermore, we formulate relative utility and relative loss functions within the optimistic three-state 3WD model, grounded in Prospect-Regret theory, and implement the model using Python. We also examine the threshold characteristics arising from psychologically perceived values in the optimistic three-state 3WD framework. In addition, we present a classification method for alternatives using the three-state 3WD model, with its implementation detailed in Python. To demonstrate the feasibility and practical value of the proposed approach, we apply it to the problem of evaluating the impact of interactive learning for blind students. The recommended methodology performs better with regard to decision-making capacity than several other methods that have been established in the existing literature.
{"title":"A fusion based optimistic three-state three-way decision framework integrating prospect-regret theory under fuzzy preference relations and their applications","authors":"Peide Liu , Abbas Ali , Noor Rehman , Areej Qadeer","doi":"10.1016/j.ins.2025.123046","DOIUrl":"10.1016/j.ins.2025.123046","url":null,"abstract":"<div><div>The three-way decision (3WD) model has gained widespread application in multi-attribute decision-making. However, existing models often neglect the variability in decision-makers minimum acceptance levels and risk attitudes across different criteria. With growing complexity and uncertainty in decision contexts, accurately capturing evaluation values remains a key challenge. In this paper, we address uncertainty in multi-attribute decision-making by introducing a novel type of fuzzy preference relation and developing collective fuzzy preference relations derived from multisource fuzzy information. To reduce the influence of subjective factors, we propose a new approach for calculating conditional probabilities based on collective fuzzy preferences and attribute weights. Furthermore, we formulate relative utility and relative loss functions within the optimistic three-state 3WD model, grounded in Prospect-Regret theory, and implement the model using Python. We also examine the threshold characteristics arising from psychologically perceived values in the optimistic three-state 3WD framework. In addition, we present a classification method for alternatives using the three-state 3WD model, with its implementation detailed in Python. To demonstrate the feasibility and practical value of the proposed approach, we apply it to the problem of evaluating the impact of interactive learning for blind students. The recommended methodology performs better with regard to decision-making capacity than several other methods that have been established in the existing literature.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"735 ","pages":"Article 123046"},"PeriodicalIF":6.8,"publicationDate":"2026-01-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145978688","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
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