Purpose Industrial predictive maintenance for ultra-high-pressure waterjet equipment faces a critical challenge: traditional survival models relying on the constant proportional hazards assumption fail to capture the nonlinear, time-varying failure mechanisms inherent to complex multi-sensor industrial systems. This paper aims to address this limitation by proposing SurvWaterJet, a novel deep survival analysis framework tailored for the web-enabled waterjet predictive maintenance system, enabling real-time failure risk prediction and remote monitoring for Industrial Internet of Things (IIoT)-connected waterjet cutting equipment. Design/methodology/approach SurvWaterJet integrates three complementary components: (1) a nuclear norm imputation module that exploits the low-rank structure of device feature matrices to recover missing sensor values with theoretical convergence guarantees; (2) a Parallel-Encoder with multi-head self-attention that explicitly models dynamic coupling among nine heterogeneous physical sensor signals; and (3) a Kolmogorov–Arnold Network sub-network with learnable B-spline activations, combined with a Piecewise Constant Hazard loss function, to approximate highly nonlinear lifecycle degradation trajectories. Findings Evaluated on a real-world WaterJet Dataset collected from 15 factory-deployed devices across 30 independent runs, SurvWaterJet achieves a Concordance Index of 0.738 (+18.3% over CPH), an integrated Brier score of 0.108 and a peak time-dependent AUC of 0.765 at the high-load operating phase (t = 1200 h), outperforming representative baselines (e.g. DeepSurv, DSM, SurvTrace, etc.) across all six evaluation metrics. The deployable RESTful API enables real-time failure risk prediction with sub-100 ms latency for up to 50 concurrent IIoT-connected devices. Originality/value To the best of the authors’ knowledge, this study presents the first application of Kolmogorov–Arnold Networks within a deep survival analysis framework for industrial equipment maintenance, specifically optimized for a web-enabled waterjet predictive maintenance system. We demonstrate that learnable univariate function compositions provide superior approximation of nonlinear degradation phenomena compared to conventional fixed-activation deep models, while the end-to-end framework offers a principled, web-deployable solution for lifecycle-aware maintenance decision-making in IIoT environments. The proposed approach bridges advanced survival modeling with web-based operational intelligence, facilitating remote monitoring and proactive intervention for waterjet cutting equipment.
PurposeThis study aims to look at how important the Internet of Things (IoT) is to creating smart cities. It emphasizes how IoT technologies maximize the use of resources, improve public services and raise the standard of living for citizens. Furthermore, it investigates how IoT may support sustainability while addressing important urban issues including energy use, transportation congestion and public safety. Design/methodology/approachThis study explores the IoTs’ architecture, uses and difficulties in smart cities. The study highlights the use of data analytics techniques to enable the integration of connected devices with cutting-edge technology to address problems like scalability, security and data overload. Particularly in relation to IoT-enabled smart city infrastructure and services, the function of data analytics in converting unprocessed IoT data into useful insights is being explored. FindingsThe report emphasizes how IoT is essential to building smart cities because it promotes efficiency, adaptability and urban environments that are focused on the needs of their citizens. Real-time interaction, optimizing resources and urban oversight are all enhanced by IoT-enabled smart city systems. Real-time data gathering, analysis and visualization are made possible by Web platforms, portals and applications, which are essential for integrating IoT data for smart city management. Key urban indicators like traffic flow, quality of air and energy usage can be tracked by municipal authorities using interactive interfaces offered by Web-based IoT dashboards. Urban management is improved by real-time online monitoring systems, which track smart city infrastructure continually and optimize waste collection, traffic lights and public safety responses. Furthermore, large-scale data processing, storage and predictive analytics are made possible by cloud-based Web services like Microsoft Azure IoT and AWS IoT, which guarantee effective resource allocation and decision-making. Smart cities may develop data-driven, adaptable urban settings, increase citizen services and increase sustainability by using these technologies. Originality/valueThe study thoroughly examines how IoT affects smart cities, highlighting how it can address urban issues while promoting sustainability. IoT and modern data analytics together reveal ideas that improve cities’ responsiveness to citizen demands, efficiency and flexibility. This study contributes by providing information on cutting-edge strategies that aid in the creation of robust and citizen-focused smart urban infrastructures.
Purpose This paper aims to address the limitations of existing aspect-based sentiment analysis (ABSA) approaches. ABSA is a fine-grained sentiment analysis task that is designed to predict the sentiment polarity of a given aspect within a sentence. Most existing approaches focus on modeling a single semantic or syntactic structure, failing to achieve comprehensive integration of multi-source information. In particular, positional information is often inadequately modeled, thereby limiting the ability to capture spatial relationships between opinion words and target aspects. Design/methodology/approach A novel tri-graph convolutional network, referred to as the Semantic, Emotional and Positional Graph Convolutional Network (SEPGCN), is proposed to address these limitations. The framework consists of a semantic graph convolutional network, an emotion-enhanced graph convolutional network and a position-enhanced graph convolutional network, which extract aspect-related features from semantic, emotional and positional perspectives. A dependency-centrality-based position fusion module is introduced to strengthen spatial awareness. Additionally, a Tri-Affine interaction module is used to facilitate deep interaction and fusion among heterogeneous features. Findings Experimental results obtained on three benchmark data sets indicate that SEPGCN outperforms representative baseline models in terms of accuracy and F1 score. In particular, it achieves up to 0.48% improvement in Accuracy and 1.22% improvement in F1 score over the strongest baselines across the datasets, while also yielding consistent gains in performance. Originality/value This study presents a tri-graph framework that jointly models semantic, emotional and positional information within a unified architecture. The findings highlight the role of positional modeling in capturing spatial associations between aspects and opinion expressions.
Purpose This study aims to address the problems of data provenance, trust management and reliable knowledge inference in Web-based biomedical information systems by developing a blockchain-enabled heterogeneous graph learning framework for prostate cancer driver gene prediction.Design/methodology/approach A framework named BE-HGAT is proposed to integrate blockchain-based provenance verification with trust-aware heterogeneous graph attention learning. Processed gene expression profiles and protein-protein interaction metadata are hashed, organized through Merkle trees and anchored to an Ethereum smart contract, while raw biomedical data remain off-chain. The resulting verification signals are combined with empirical interaction confidence scores to construct trust-weighted edges in a heterogeneous information network. Node-level and semantic-level attention mechanisms are then used to learn representations across genes, samples and disease-related semantic paths.Findings Experiments on publicly available prostate cancer data sets show that BE-HGAT outperforms representative graph learning baselines in Accuracy, Precision, Recall, F1-score and ROC-AUC. The model achieves an average F1-score of 0.777 and a ROC-AUC of 0.938, exceeding the strongest baseline by approximately 0.075 in F1-score and 0.04 in ROC-AUC. Robustness and ablation analyses further indicate that heterogeneous semantic attention provides the main performance gain, while blockchain-derived trust signals improve stability under noisy and perturbed graph conditions.Originality/value This study links decentralized data provenance with heterogeneous graph-based biomedical inference. Rather than using blockchain only as a storage or audit mechanism, BE-HGAT converts cryptographic verification outcomes into trust-aware graph signals, enabling more traceable, robust and reliability-aware driver gene prediction in distributed biomedical information systems.
Purpose This study aims to address the dual challenges of source-domain annotation subjectivity and target-domain pseudo-label noise in cross-domain Aspect Sentiment Triplet Extraction (ASTE). The purpose is to enhance model reliability and robustness under domain shift, thereby supporting fine-grained sentiment understanding and trustworthy decision-making for intelligent agents in web information systems (WIS). Design/methodology/approach Adopting a "text-label-text" paradigm, transductive data augmentation framework for cross - domain ASTE using large language models (TDA-LLM) systematically orchestrates large language models (LLMs) as controllable augmentation agents across three stages: source-domain annotation refinement to reduce label ambiguity, target-domain pseudo-data generation to bridge domain gaps and prediction post-processing with structural validation. A bidirectional consistency-based filtering mechanism prunes unreliable samples, followed by transductive inference for dependency-aware validation. Findings Extensive experiments on six cross-domain benchmarks show that TDA-LLM outperforms strong state-of-the-art baselines by an average of 1.78% in F1 score. The framework effectively mitigates annotation inconsistency and pseudo-label noise, improving structural coherence and prediction reliability in cross-domain sentiment triplet extraction. Originality/value The originality of TDA-LLM lies in its systematic integration of LLMs as controllable augmentation agents throughout the entire data pipeline - from source-domain annotation refinement and target-domain pseudo-label generation to transductive inference with structural validation. It introduces a novel bidirectional consistency-based filtering mechanism to ensure semantic and syntactic fidelity. The value resides in its demonstrated ability to robustly mitigate cross-domain supervision noise, thereby advancing the reliability and trustworthiness of fine-grained sentiment understanding systems for intelligent agents in WISs.
Purpose A two-layer stacked LSTM architecture is integrated with multi-head temporal attention (eight heads) to capture short- and long-term degradation patterns, and a feature attention module to dynamically weight sensor channels. An asymmetric loss function penalizes overestimation, and attention regularization promotes head diversity. The model is trained and validated on the NASA commercial modular aero-propulsion system simulation (C-MAPSS) data set using RMSE, R1 and custom metrics. Design/methodology/approach This study aims to enhance the accuracy and reliability of Remaining Useful Life (RUL) prediction for aircraft engines in Prognostics and Health Management (PHM) systems. By addressing limitations in standard LSTM models, such as capturing long-range dependencies and handling noisy sensor data, the research proposes a modified LSTM framework tailored for multivariate time-series data from IoT-enabled engines in Web-based environments. Findings Experiments on C-MAPSS subsets demonstrate superior performance: achieving an RMSE of 17.80 compared to 28.21 for the baseline LSTM. Multi-head attention (eight heads) outperforms variants (2 / 4/16 heads), balancing complexity and accuracy, especially under asymmetric penalties. Research limitations/implications Limited to C-MAPSS data set; real-world deployment may require adaptation to diverse fault modes. Implications include advancing attention-enhanced models for non-stationary signals, with future extensions to federated learning and domain-specific integrations for broader PHM applications. Practical implications This study provides significant practical contributions to the field of PHM within Industrial IoT (IIoT) frameworks. By achieving high-precision RUL predictions (with RMSE reduced to the 19.90–28.93 range), the proposed model offers a robust technical foundation for transitioning from traditional scheduled maintenance to cost-effective predictive maintenance, maximizing component utility while minimizing unscheduled downtime. Social implications Enhances flight safety by enabling proactive interventions, minimizing unexpected failures. Promotes sustainable aviation through extended component lifespans, reducing environmental impact from frequent replacements and supporting global safety standards. Originality/value Novel hybrid LSTM with hierarchical multi-head temporal and feature attention, plus asymmetric loss and regularization, outperforms state-of-the-art methods in accuracy (up to 60\% RMSE reduction) and reliability.
PurposeThis paper aims to address modality noise and continuous changing of reliabilities of emotional cues in multimodal sentiment analysis. Such dynamic perceptual characteristics are challenging for continual sentiment perception, especially when there are modality conflicts.Design/methodology/approachThe authors propose a Continuously self-Adaptive Multimodal Sentiment Analysis method (CAMSA). CAMSA decouples modality features into shared and private features and introduces a Variational Information Bottleneck method to perform probabilistic modeling and information compression on shared features, and then continually suppressing redundant and noisy information during the progressively emerging of emotional cues. In addition, a multidimensional adaptive feature fusion mechanism is designed to evaluate modality reliability from the perspective of semantic certainty, modal consistency and sentiment extremity.FindingsExperimental results on CMU-MOSI, CMU-MOSEI and CH-SIMS demonstrate that CAMSA outperforms existing models on most evaluation metrics, validating its effectiveness in handling modality noise and modal conflicts.Originality/valueFrom the requirements of continual sentiment perception, this paper proposes a continuously self-adaptive CAMSA framework, which combines shared feature refinement with a multidimensional adaptive fusion mechanism, enabling robust cross-modal sentiment modeling in dynamic and evolving perceptual scenarios. The CAMSA approach can effectively address the problem of fixed fusion strategies in complex dynamic sentiment perception tasks.
Purpose Text adversarial techniques are a key research area in natural language processing (NLP). With the widespread application of deep learning-driven NLP in sentiment analysis and machine translation, the lack of robustness against adversarial attacks has become increasingly evident. Adversarial samples in the text domain can mislead models and the discrete nature of text distinguishes adversarial techniques in this field from others. To outline recent progress, this paper aims to review relevant studies from the past decade, summarizing advancements in adversarial attack, defense and detection methods. Design/methodology/approach This paper conducts a systematic literature review of text adversarial attacks and defenses. The review begins with an analysis of attack methods, then discusses text adversarial defense and detection methods and finally points out the challenges in both attack and defense. Findings The authors provide a clear classification of attack methods based on specific criteria, followed by an overview of defense and detection techniques, highlighting their respective strengths and limitations. Originality/value This review systematically classifies and analyzes text adversarial attack, defense and detection techniques and thoroughly explores the challenges and future development directions in this field, providing valuable reference information for researchers in this area.
PurposeThis paper aims to address several limitations of current graph neural network (GNN)-based bundle recommender systems. These models often suffer from representation degradation as the number of aggregation layers increases and neglect latent interaction patterns that are essential for improving recommendation performance. To address these limitations, we propose a novel model named hierarchical graph-based mixup for bundle recommendation (HGMBR), which integrates mixup techniques across both the item and bundle views to learn higher-quality representations for recommendation. Design/methodology/approachThe proposed model, HGMBR, uses a multilayer perceptron (MLP)-based residual network to extract latent interaction features from user–bundle and user–item embeddings. A graph-based mixup module then performs interpolation within both the bundle and item views. This process effectively enriches the user and bundle representations by capturing both explicit and latent interactions. In addition, a cross-view mixup module aligns user representations across the two views. FindingsExtensive experiments on three publicly available data sets demonstrate that HGMBR alleviates representation degradation in deep aggregation layers and mitigates bundle sparsity issues, demonstrating its superiority over existing models in bundle recommendation tasks. Originality/value HGMBR captures latent interaction features and uses mixup techniques hierarchically to enrich user and bundle representations. This results in robust, discriminative embeddings that are less susceptible to noise and significantly enhance recommendation performance.
Purpose - Multilingual social networks present significant challenges in malicious group detection due to linguistic diversity and heterogeneous data streams amplified by artificial intelligence (AI) and Internet of Things (IoT) technologies. This study aims to propose a novel framework to enhance detection robustness, accuracy and sensitivity in these environments. Design/methodology/approach - Leveraging graph representation learning (GRL), the solution integrates multilingual natural language processing, machine learning and structural community discovery. The framework incorporates multimodal data (including user positional information) to optimize real-time monitoring and early warning systems. Findings - Quantitative evaluation demonstrates state-of-the-art performance: 95% classification accuracy and 94.88% macro-F1 score for group influence recognition, representing statistically significant improvements of 2.5% (accuracy) and 2.44% (F1) over baselines (p < 0.01). Under high-confidence constraints with positional data, performance increases to 99.17% accuracy and 98.33% macro-F1. Originality/value - This work addresses critical gaps in cross-lingual threat detection by unifying GRL with multimodal AI/IoT data processing. The proposed framework advances real-time malicious group identification in Web information systems, offering scalable solutions for security-sensitive platforms.
PurposeThis paper aims to enhance the quality of images generated by electroencephalography (EEG) encoding and decoding models; this paper improves the neurological plausibility of the generated images by emulating key mechanisms of human visual processing.Design/methodology/approachMotivated by the center-periphery organization of visual perception, an EEG segmentation module and a feature fusion module are introduced into a diffusion-based EEG-to-image generation framework. The EEG segmentation module decomposes the input signal into a segment with high fluctuation amplitude and the remaining signal, which are separately encoded and subsequently fused to enable differentiated visual modeling. In addition, a neuro-inspired application framework is proposed to extend the EEG-to-image generation approach to Web search scenarios, where EEG-generated images serve as implicit visual representations of user intent.FindingsExperimental results demonstrate that integrating EEG segmentation and feature fusion leads to measurable improvements in the perceptual quality and structural coherence of EEG-generated images.Originality/valueAn EEG segmentation module and a feature fusion module are introduced into a diffusion-based EEG-to-image generation framework. The EEG segmentation module decomposes the input signal into a segment with high fluctuation amplitude and the remaining signal, which are separately encoded and subsequently fused to enable differentiated visual modeling. A neuro-inspired application framework is proposed to extend the EEG-to-image generation approach to Web search scenarios, where EEG-generated images serve as implicit visual representations of user intent.
PurposeThis study aims to investigate the impact of complex task dependencies and dynamic vehicular environments on task offloading efficiency in vehicular edge computing (VEC) systems, addressing scalable Quality of Service-aware scheduling for dependency-intensive Web services. It seeks new insights into minimizing latency and energy consumption while meeting service-level agreement (SLA) requirements, advancing understanding of resource management in intelligent vehicular networks.Design/methodology/approachThis study proposes a dependency-aware Web-related task offloading framework that models interrelated tasks as directed acyclic graphs (DAGs). Data from simulated VEC scenarios - incorporating real-time vehicle mobility patterns, SLA-defined task priorities and road traffic density predictions - were analyzed through Python simulations to evaluate the proposed dynamic association particle swarm optimization (DAPSO) algorithm against benchmark schedule algorithms.Findings The results demonstrate that the DAPSO framework effectively reduces task offloading latency and energy consumption through predictive edge server resource reservation. This empirically validates the critical necessity of integrating dependency-aware heuristic algorithms with proactive resource allocation mechanisms to address NP-hard scheduling challenges in dynamic VEC environments.Originality/valueBy integrating DAG-based dependency modeling, real-time mobility awareness, SLA prioritization and predictive resource reservation into a unified VEC framework, this research provides theoretical and practical foundations for next-generation intelligent transportation systems. The DAPSO algorithm offers implementable solutions for latency-sensitive Internet of Vehicle applications while highlighting pathways for adaptive large-scale optimization.
PurposeThis paper aims to address the challenges of dynamic and web-oriented network resource provisioning in the Internet of Vehicles (IoV), where the diversity and complexity of in-vehicle web services require efficient and adaptive resource scheduling. It proposes a dual-time-scale strategy to enhance resource allocation, reduce service scheduling delays and ensure load balancing across network slices.Design/methodology/approachThis study adopts a dual-time-scale approach for network slicing resource prediction and dynamic scheduling. At the long-time scale, a long short-term memory (LSTM) network is employed to learn temporal patterns of vehicle behavior and web service requests, enabling accurate load prediction and proactive resource preconfiguration. At the short-time scale, a dueling double deep Q-network (D3QN) algorithm performs fine-grained resource scheduling based on real-time service states, ensuring compliance with service level agreements. Experiments are conducted to evaluate the strategy's performance in terms of scheduling latency, system utility and load balancing.FindingsExperimental results demonstrate that the proposed LSTM-D3QN strategy maintains high system utility saturation while significantly reducing web service scheduling latency during peak periods. It also effectively ensures load balancing among slices. Compared to baseline algorithms, the method achieves an average reduction in service scheduling delay by 54.57%, an improvement in utility saturation by 22.41% and a decrease in slice load by 23.46% across three typical service slices.Originality/valueThis paper introduces a novel dual-time-scale framework that integrates LSTM-based load prediction with D3QN-based real-time scheduling for web-oriented IoV environments. By jointly optimizing long-term resource preallocation and short-term dynamic adjustment, the approach enhances adaptability and robustness in dynamic service provisioning. The study provides a comprehensive solution to the challenges of heterogeneous web service demands in IoV, offering significant practical and theoretical insights for future network slicing and resource management research.
PurposeThe purpose of this paper is to introduce Nuwa, a robust fuzzing framework for testing real-world Windows applications (UWP, WinUI2, WinUI3 and Win32), addressing challenges like multi-process architectures and complex graphical user interfaces (GUIs) that hinder existing tools.Design/methodology/approachNuwa integrates: a lightweight instrumentation module using WinDbg C++ API for code coverage; a scheduling algorithm prioritizing inputs based on coverage feedback; and an automated interaction layer leveraging Windows automation APIs for GUI manipulation with malformed inputs. The instrumentation module provides real-time coverage feedback, which is used by the scheduling algorithm to prioritize inputs that are more likely to trigger new code paths. The automated interaction module ensures that these inputs can be delivered to the application, even when complex GUI interactions are required.FindingsTesting on Siemens Solid Edge Free Viewer and Windows Media Player (UWP) revealed critical vulnerabilities, including out-of-bound writes, reads and zero day. Siemens yielded 9 CVEs (CVE-2023-39181 to CVE-2023-39186, CVE-2024-47941, CVE-2024-47940 and CVE-2024-54091), while Media Player had 3 undisclosed zero day (2 AVI, 1 MOV parsing), impacting millions of devices and causing Windows Explorer crashes. Experiments show Nuwa also outperforms WinAFL in both DynamoRIO and Intel-PT modes in metrics including compatibility, fuzzing throughput and coverage.Originality/valueTo the best of the authors' knowledge, Nuwa is the first coverage-guided fuzzing framework for real-world Windows apps, combining efficient instrumentation, smart scheduling and automated GUI interaction to fill a critical gap in fuzzing tools.
PurposeIn time-critical natural disaster scenarios, unmanned aerial vehicles (UAVs) are crucial for search and rescue. While mobile edge computing (MEC) enables real-time data processing for these UAVs, it introduces a significant challenge: balancing low-delay data analysis to locate survivors against the UAVs' limited battery life. This paper aims to propose a solution to minimize task processing delay in dynamic rescue environments while conserving UAV energy.Design/methodology/approachTo overcome this challenge, this study proposes the multi-queue Lyapunov-guided deep reinforcement learning (MQ-LyDRL) method to minimize task processing delay by jointly optimizing task offloading and resource allocation. This method innovatively integrates Lyapunov optimization with DRL. Specifically, by constructing Lyapunov functions based on queue stability and energy constraints, MQ-LyDRL decomposes the complex multistage stochastic optimization problem into a deterministic, per-time-slot subproblem. An adaptive DRL framework is then employed to solve this subproblem, enabling it to learn the optimal policy for real-time decision-making without requiring prior knowledge of the environment's dynamics.FindingsExtensive simulations demonstrate that MQ-LyDRL significantly outperforms existing methods. It maintains operational stability in fluctuating conditions and reduces average delay by at least 9.21% while adhering to an energy budget. This reduction translates to faster data-to-decision cycles, accelerating life-saving interventions by extending the operational time of UAVs.Originality/valueThis work's primary value is providing a blueprint for intelligent and efficient edge computing systems in high-stakes scenarios. By combining stability theory with adaptive artificial intelligence (AI), this study offers a practical framework applicable to critical missions where performance and reliability are nonnegotiable.
Purpose The purpose of this study is to enhance AI model explainability and value alignment in resource-limited environments by transforming the traditionally opaque black-box models into transparent, interpretable systems. By embedding DIKWP semantic reasoning and System 2 cognitive control into distributed learning frameworks, this study seeks to monitor and guide inference paths in real-time. This ensures alignment with user purposes and security expectations without incurring high resource costs, offering a viable solution for privacy-preserving, trustworthy AI deployment on edge devices such as smartphones, wearables and home IoT systems. Design/methodology/approach This study proposes a DIKWP-based white-box semantic distributed learning framework tailored for resource-constrained devices. It integrates dual-process cognitive theory (System 1/System 2) and embeds semantic probes into models to monitor DIKWP transformations—Data, Information, Knowledge, Wisdom, Purpose—during inference. A DIKWP×DIKWP transformation matrix quantifies semantic transitions, enabling transparent reasoning path tracking. Lightweight probe mechanisms allow model introspection with minimal computational overhead. The framework is evaluated on reasoning-intensive data sets via metrics such as semantic unit coverage, cognitive path entropy and reasoning step frequency, validating its effectiveness in enhancing explainability and safety under federated learning and bandwidth-constrained environments. Findings Experimental results across CMMLU, Math23K and MMLU data sets show that the DIKWP-WISE framework significantly improves semantic reasoning depth, transformation coverage and cognitive entropy compared to traditional models. Models using DIKWP probes exhibit higher ratios of System 2 (deliberative) reasoning, fewer inappropriate responses and better purpose alignment. Notably, even under resource constraints, the semantic probes maintain performance without adding substantial computational load. Moreover, the framework enables semantic-level federated reasoning, contributing to both model safety and explainability, particularly in knowledge-intensive or user-critical tasks such as education, health care and intelligent interaction. Originality/value This paper pioneers a semantic white-box reasoning framework combining the DIKWP model with cognitive psychology to achieve real-time introspection of AI models in edge environments. Unlike traditional post-hoc explainability techniques, the DIKWP-WISE architecture embeds transparent reasoning directly into the model’s operation using semantic probes. This approach uniquely aligns semantic understanding with purpose-driven inference and provides a scalable, architecture-agnostic mechanism for secure, explainable AI on low-power devices. It bridges the gap between symbolic and sub-symbolic reasoning while offering practical contributions to secure federated learning and human-aligned decision-making systems.
Purpose The purpose of this study is to address the task of visually rich document understanding, which extracts structured semantics from complex multimodal document images. Core challenges, including feature distortion from geometric deformations and structural deviations from complex layouts, jointly undermine extraction precision and robustness – the former due to inadequate cross-modal adaptation to local distortions and the latter due to constrained global topology modeling. Design/methodology/approach To address these limitations, the Geometry-Topology Collaborative Parsing Framework is proposed. This framework achieves robust document parsing through dual technical approaches. First, deformable kernels are used to correct geometrically warped text features, with channel attention mechanisms integrated for feature noise suppression. Second, cross-attention mechanisms and the XY-Cut algorithm work collaboratively to construct reading orders adapted to document topological structures. This co-design ensures simultaneous optimization of text feature enhancement and reading-order construction. Findings The framework demonstrates superior performance in document classification and structured information extraction through synergistic optimization of geometric correction and structural relationship decoding, significantly reducing reliance on error-prone optical character recognition (OCR) preprocessing. Originality/value This work establishes a novel paradigm for joint spatial-semantic document understanding via three key advancements: adaptive distortion rectification through deformable-convolutional text localization; density-aware reading sequence generation for robust layout parsing; and unified representation learning bridging visual features with structural topology without intermediate OCR dependencies.
PurposeThis study tackles the critical gap where existing cloud service selection methods fail small and medium manufacturers (SMMEs). This study aims to overcome the limitations of overlooking SMMEs' unique resource constraints, acute risk sensitivity and diverse behavioral preferences. The core purpose is to provide SMMEs with a scientifically grounded, interpretable and operationally feasible framework that genuinely captures their authentic needs and risk perceptions, enabling them to strategically select cloud services that maximize operational benefits and mitigate adoption risks.Design/methodology/approachThis study developed a novel behavioral framework combining Kano-based demand classification, regret theory for dynamic weighting and a prospect-evaluation based on distance from average solution (EDAS) model. A multi-layered evaluation system (functionality, tech, cost, supplier, service) was built. Kano questionnaires refined demands and mapped regret sensitivities. Membership degrees were dynamically adjusted using a p,q-order orthogonal dual hesitant fuzzy matrix. Critically, prospect theory reference points integrated gain-loss perceptions into the final ranking.FindingsEmpirical testing with an SMME confirmed the framework's effectiveness. It successfully captured authentic SMME preferences regarding tech needs, cost sensitivity and service expectations. The integrated behavioral approach offered superior interpretability and practical feasibility compared to traditional static or purely subjective methods.Originality/valueThis study presents a novel integration of Kano analysis, regret theory and prospect theory within an EDAS framework, specifically tailored for SMMEs. The key methodological innovations involve: i) dynamic mapping between demand attributes and regret sensitivity coefficients, and ii) incorporation of behavioral economic principles (gain-loss perception) into service ranking. The resulting behaviorally-augmented decision tool enables SMMEs to achieve risk-optimized cloud adoption through needs-based service alignment.
Purpose The rank aggregation problem, which has many real-world applications, refers to combining multiple input rankings into a single aggregated ranking. In dynamic settings, where new rankings arrive over time, efficiently updating the aggregated ranking is essential. This paper aims to develop fast, theoretically grounded and practically efficient algorithms for dynamic rank aggregation. Design/methodology/approach The authors first develop left right (LR) aggregation, built on the LR tree data structure. The LR tree is inspired by the LR distance, a novel but equivalent formulation of the classical Spearman’s footrule distance, designed to support efficient incremental updates. They then analyze the classical Pick-A-Perm algorithm under Spearman’s footrule distance and show how it can also be maintained efficiently in the dynamic setting. Finally, they combine LR aggregation and Pick-A-Perm into a unified dynamic rank aggregation framework that returns the better of the two candidate aggregations at each step. Findings Experimental evaluations show that LR aggregation produces solutions close to optimal in practice. They prove that Pick-A-Perm yields an expected 2-approximation under Spearman’s footrule distance and they show that both LR aggregation and Pick-A-Perm (as well as their combination) can be implemented with O (n log n) update time and O(n2) space, independent of the number of rankings received. Originality/value To the best of the authors’ knowledge, this work provides the first near-linear-time dynamic rank aggregation framework that offers both a provable approximation guarantee and strong empirical performance in practice.
