Search Results (3,553)

Geographical knowledge graphs (GeoKGs) have long experienced several fundamental challenges in representing complex spatial relationships, such as limited dimensionality, insufficient quantification of relationship strength, and weak reasoning capabilities. To address these issues, this study presents the multidimensional spatial relation knowledge graph (MDSR-KG) framework. The novelty of this framework lies in advancing the shift toward spatial relation node-based representation, thereby elevating the spatial relations from edge structures to independent, computable, and inferable structured nodes. This approach was complemented by a parametric method aimed at quantifying the relation strength between nodes, thereby facilitating an advancement from discrete relations to continuous and interpretable association weighting. In experiments conducted in this study using the Berlin OpenStreetMap data, we noted that for complex spatial queries, the MDSR-KG framework significantly outperformed the baseline models in accuracy and completeness. The framework also exhibited advanced reasoning capabilities, such as ranking and recommendation, which are lacking in traditional methods. Thus, the framework lays a theoretical foundation for advancing from geographic feature recognition to spatial relationship comprehension. Full article

Securing IoT-enabled wireless sensor network links requires selecting a minimum-cost set of connected monitor nodes that observes every link while satisfying capacity constraints, a problem known as the minimum weighted connected capacitated vertex cover (MWCCVC). To the best of our knowledge, this work introduces the first learning-based framework for the MWCCVC through a three-stage pipeline that combines supervised graph neural networks, feasibility repair, and local search. We compare twelve graph neural network architectures, including graph convolutional network, graph attention network, GraphSAGE, Graph Isomorphism Network (GIN), and GraphTransformer, under unified features, loss functions, and hyperparameter tuning. Throughout the evaluation on 309 benchmark instances under a 5-fold cross-validation protocol, feasibility is guaranteed by the deterministic repair module instead of being learned by the network, resulting in 100% feasible covers across all evaluated instances. At the large scale, GIN, GraphSAGE, DeeperGIN, and EdgeAwareGIN reach parity with the state-of-the-art hybrid genetic algorithm (HGA), with GIN attaining a mean gap of $-0.37\%$ (a difference of less than one percentage point) while completing in seconds instead of HGA’s hours. Statistical tests across the full 309-instance benchmark confirm significant differences between the architectures, with Friedman ${\chi }^2=93.05$, $p<{10}^{-4}$. The best-performing architectures remain within about $2\%$ of HGA on small- and medium-scale instances, where HGA is near-optimal, and become the preferred choice at the large scale, mainly because their wall-clock time is much shorter than HGA’s at the same solution quality. Full article

Open-set recognition of wild plants in natural complex scenes is an important task for plant conservation, ecological monitoring, and precision agriculture. Traditional closed-set learning methods struggle to handle unseen species not covered by the training set and complex environmental interferences, while existing open-vocabulary methods lack knowledge-driven reasoning capabilities and cannot provide interpretable recognition for unknown categories. This research proposes the Reasoning-Aware Perceptual Framework that integrates open-vocabulary vision-language models, foundation mask-generation tools, and domain knowledge reasoning to achieve known/unknown category recognition, online perception, and interpretable reasoning of unknown wild plant species. Centered on a five-stage closed loop of Perception-Retrieval-Reasoning-Decision-Iteration, the framework captures open concepts through vision-language feature alignment, completes evidence-based reasoning and confidence evaluation in combination with a botanical domain knowledge base, and finally outputs species classification decisions, interpretable reasoning reports with family/genus-level taxonomic affinity, and uncertainty-calibrated confidence scores. The unknown category estimation with family/genus-level taxonomic affinity in this framework refers to a general unknown label combined with taxonomic affinity at the family/genus level, which can clearly reflect the evolutionary relationship between unknown species and known species. Experiments on the self-constructed WildPlantOpenSet-10K dataset and public benchmark datasets report an F1-score of 84.7% for unknown species recognition, AUROC of 0.93 for known/unknown discriminability, and mean F1 of 87.0% across all categories. This framework focuses on open-set wild plant recognition and interpretable reasoning, using off-the-shelf instance extraction to acquire visual features for downstream reasoning. It maintains stable robustness in complex scenarios such as occlusion, strong light, and multi-species coexistence, and can adapt to the open-world environment without relying on large-scale pixel annotations, providing a research prototype for interpretable open-set recognition in complex natural environments. Full article

Securing the Industrial Internet of Things (IIoT) is paramount, yet proprietary protocols remain vulnerable to deep-state logic flaws that traditional fuzzers often fail to reach. We propose MALF, a Multi-Agent LLM Fuzzing Framework that couples a dynamic Industrial Security Knowledge Graph (ISKG) with collaborative cognitive agents for effective, efficient, and trustworthy IIoT security testing. A self-evolving knowledge loop mitigates LLM hallucinations by grounding the generation in verifiable graph constraints; QLoRA-tuned models aligned with hexadecimal features enable low-latency mutation; and Chain-of-Thought reasoning reconstructs protocol states for intent-driven attacks. On a heterogeneous testbed spanning five industrial protocols and ten vendors, MALF achieves an average Test Case Acceptance Rate of 88.3% (peak 91.2% on Modbus/TCP) and 91.2% ISKG-defined state coverage, outperforming rule-based, RL-based, and LLM baselines. On a 15-vulnerability N-Day benchmark, MALF detects all known cases, against 60%, 47%, 40%, and 27% for NCMFuzzer, MARLFuzz, BooFuzz, and Fuzz4All, respectively. In a separate real-world campaign, MALF further identifies 14 previously unknown vulnerability candidates, of which four have been assigned CNVD identifiers (CNVD-2024-16009, CNVD-2025-22875, CNVD-2025-29811, CNVD-2026-06041) and 10 remain under vendor review. These results provide controlled-testbed evidence that knowledge-grounded AI agents can systematically expose deep-state vulnerabilities in opaque IIoT environments. Full article

Accurate and automatic segmentation of the liver and liver tumors from computed tomography (CT) images is essential for computer-assisted diagnosis, treatment planning, and clinical decision-making. Although deep learning-based segmentation models, particularly U-Net and its variants, have achieved promising results in medical image analysis, many existing approaches mainly focus on local pixel-level feature extraction and may have limited ability to explicitly model long-range spatial relationships among anatomically meaningful regions. In addition, liver tumor segmentation remains challenging due to low contrast, irregular tumor boundaries, heterogeneous tumor appearances, and noise or artifacts in CT images. To address these limitations, this study proposes a hybrid ensemble neural network architecture that integrates an improved U-Net and a Graph U-Net for automatic liver and liver tumor segmentation. The improved U-Net is designed to capture fine-grained local features and preserve detailed spatial information through an encoder–decoder structure with skip connections, while the Graph U-Net uses Simple Linear Iterative Clustering (SLIC)-based superpixels to construct a graph representation of CT images and model spatial dependencies between adjacent image regions. By combining these complementary representations through an ensemble learning strategy, the proposed framework enhances both pixel-level segmentation accuracy and robustness against noisy imaging conditions. The proposed method was evaluated on the LiTS17 dataset, where CT images were preprocessed using intensity filtering, resizing, data augmentation, and normalization. Experimental results demonstrate that the proposed ensemble architecture achieves 99.2% accuracy for liver segmentation and 98.1% accuracy for liver tumor segmentation, outperforming representative segmentation models such as MultiresUnet and R2U-Net. Furthermore, robustness experiments under different signal-to-noise ratio conditions show that the proposed model maintains stable performance in noisy CT images, achieving 85% accuracy even under severe noise at −4 dB SNR. This result highlights the advantage of integrating convolutional feature learning with graph-based spatial relationship modeling for improving segmentation stability when image quality is degraded by noise or artifacts. These findings indicate that the integration of improved U-Net, SLIC-based graph construction, and Graph U-Net provides an effective and noise-robust solution for liver and liver tumor segmentation, with potential applicability as a computer-assisted tool in clinical image analysis after further validation on larger and external datasets. Full article

Bridge structural health monitoring (SHM) systems often suffer from large-scale data missing due to sensor faults, communication interruptions and other reasons during long-term operation, which seriously restricts the reliability of structural state assessment and maintenance decision-making. Compared with conventional single-channel independent modeling strategies commonly used for data imputation, their inherent neglect of spatial correlations and cross-modal causal associations among multi-source heterogeneous monitoring data such as displacement, wind speed, and temperature constrain the imputation capability, particularly when the target channel suffers from long-term continuous data loss. To address the above problems, this paper proposes a collaborative imputation framework integrating a graph attention network (GAT), a modal-aware cross-attention (MACA) mechanism and temporal encoder–decoder architecture (ITimeGAN). Firstly, the sensor feature topological graph is constructed based on the Pearson correlation coefficient, and the spatial dependency among multi-source features is adaptively learned through GAT. Then, the MACA module is introduced, which takes the target displacement as Query and environmental loads as Key/Value, and dynamically aggregates cross-modal driving information through multi-head attention. Finally, a bidirectional LSTM encoder and a unidirectional LSTM decoder are adopted to capture long-range temporal dependencies, so as to realize the accurate reconstruction of missing displacement data. Validated on the 9-dimensional real-world monitoring data from the GeoSHM system of the Forth Road Bridge (UK) under both random missing (10–50%) and continuous long-term missing (1–10 days) scenarios, ITimeGAN achieves an R² of 0.9950 (MAE = 4.25 mm) for longitudinal displacement and 0.9759 (MAE = 6.70 mm) for vertical displacement even under 10 consecutive days of complete data absence. Ablation analysis further reveals that the incorporation of graph attention and cross-modal attention modules reduces the longitudinal displacement MAE by 57% over the baseline, with the imputation performance ranking across three displacement directions being fully consistent with the underlying physical correlation strengths, thereby confirming the effectiveness of the proposed cross-modal collaborative strategy. Full article

Skeletal kinetic mechanisms are essential for reducing the computational cost of ammonia combustion simulations while retaining the key chemical features governing ignition, flame propagation, and NO formation. This study extends the DRG-CSP-ANN reduction and optimization framework to ammonia combustion over a broader multi-condition parameter space, aiming to develop a compact skeletal mechanism applicable to different pressures, equivalence ratios, and temperatures. Sixteen detailed ammonia combustion mechanisms were first assessed against experimental data covering ignition delay time, laminar flame speed, and NOx species concentrations over wide ranges of pressure, temperature, equivalence ratio, and oxidizer composition. Based on the overall error evaluation, the detailed mechanism with the most balanced predictive performance was selected as the parent mechanism. The parent mechanism was then reduced using the Directed Relation Graph and Computational Singular Perturbation methods, yielding an initial skeletal mechanism, RA-Ori, with 20 species and 76 reactions. To compensate for the accuracy loss caused by mechanism reduction, an Artificial Neural Network surrogate was constructed to optimize the pre-exponential factors of selected sensitive reactions within their evaluated uncertainty ranges, leading to the final mechanism, RA-ANN. The validation results show that RA-ANN reasonably reproduces ignition delay times, laminar flame speeds, and NO concentrations under different ammonia combustion conditions. Quantitatively, RA-ANN reduces the overall error from 0.335 for RA-Ori to 0.206, corresponding to a 38.4% reduction, while maintaining the same compact size. Its overall error is close to that of the parent detailed mechanism and lower than that of several existing skeletal mechanisms considered in this work. These results demonstrate that the proposed DRG-CSP-ANN strategy can construct a compact ammonia skeletal mechanism that achieves a favorable balance between computational efficiency, predictive accuracy, and applicability over representative multi-condition ammonia combustion regimes. Full article

Arabic’s complex morphological system and the optional use of short vowels (tashkīl) introduce substantial lexical ambiguity, posing significant challenges for Large Language Models (LLMs). While diacritics enhance linguistic precision, LLMs trained predominantly on undiacritized corpora often exhibit performance degradation when processing fully diacritized inputs due to representation shifts and tokenization inconsistencies. To address this limitation, we propose the Arabic Diacritic Lexical Knowledge Graph (ADL-KG), a structured framework that links diacritized and undiacritized forms through integrated lexical, morphological, and semantic knowledge. Building upon this resource, we introduce Diacritic-Aware Knowledge Graph Prompting (DA-KGP), a prompt augmentation strategy that injects explicit linguistic features into LLM inputs to facilitate robust interpretation of diacritized Arabic text. The framework is evaluated on the Arabic Reading Comprehension Dataset under zero-shot and few-shot question answering across AraGPT2-base, BLOOMZ-560M, SILMA-v1, and LLaMA 3.1-8B. Performance is assessed using Exact Match, BLEU, ROUGE-1, and BERTScore-F1. Experimental results show that fully diacritized prompts significantly degrade baseline performance, whereas DA-KGP consistently mitigates this effect by improving semantic alignment across diverse architectures. For AraGPT2-base, KG augmentation improves average BERTScore-F1 by +5.96 points. SILMA-v1 achieves the strongest lexical improvements, reaching 21.57 BLEU and 81.31% BERTScore-F1 in the KG-enhanced two-shot configuration. LLaMA 3.1-8B achieves the highest overall semantic performance with 82.54% BERTScore-F1 under KG-enhanced prompting, while BLOOMZ-560M also demonstrates statistically significant semantic gains through structured augmentation. These findings demonstrate that morphologically informed prompting and structured lexical grounding provide an effective and parameter-efficient strategy for improving the robustness and semantic fidelity of Arabic LLMs under fully diacritized input conditions. Full article

Conventional clustering techniques for urban water consumption profiling treat each household as an independent entity, thereby disregarding the spatial, socioeconomic, and infrastructural contexts that jointly govern demand behavior. This structural limitation prevents the extraction of contextually coherent consumption profiles—a critical shortcoming for utility managers who must design spatially targeted conservation interventions. To overcome this, we propose Simple GLAC, a novel graph clustering framework that leverages graph neural networks with an adaptive attention mechanism to dynamically model these complex interdependencies. The model’s end-to-end training jointly optimizes a latent representation for cluster cohesion, separation, and spatial homogeneity, where each household’s multi-month consumption record serves as the node feature vector encoding temporal consumption patterns. Evaluated on a large-scale real-world dataset of 4590 residential households across four distinct graph topologies, Simple GLAC consistently achieves superior multi-metric performance over both traditional and graph-based benchmarks, yielding interpretable and operationally actionable consumption profiles aligned with the spatial, administrative, socioeconomic, and infrastructural dimensions of urban water governance in the studied context. This work provides a data-driven tool for utility managers to deploy targeted water conservation strategies, with findings grounded in a Colombian mid-sized city and generalization to broader urban settings identified as a priority direction for future work. Full article

Motor imagery (MI)-based brain–computer interfaces (BCIs) have led to great interest as a result of their potential use in neurorehabilitation, assistive robotics, and human–computer interaction. However, decoding electroencephalographic (EEG) signals with high accuracy continues to be a difficult task due to the weak signal-to-noise ratio, differences among subjects, and the complicated temporal–spectral–spatial neural dynamics. Deep learning methods recently developed, such as convolutional neural networks, recurrent architectures, graph neural networks, and adversarial transfer learning, have enhanced MI decoding performance, yet many models are still concentrating on a single representation domain or they need costly adaptation phases in terms of computation. To tackle these shortcomings, we present NeuroCrossNet, a unified tri-modal deep learning model that is able to learn the temporal, spectral, and spatial EEG features jointly for robust and calibration-free MI decoding. The suggested network combines a Temporal HyperMixer Block for capturing long-range temporal dependencies, a wavelet transformer for learning localized time–frequency representation, and a Graph Attention Network for EEG topology-aware spatial reasoning. Additionally, a Dynamic Residual Attention Gate (DRAG) has been developed to adaptively merge heterogeneous feature streams, and a compact subject-aware normalization (SAN) method enhances cross-subject generalization without the use of labeled target-domain calibration data. Our proposed model was tested following the rigorous leave-one-subject-out (LOSO) approach on BCI Competition IV-2a and High-Gamma datasets. NeuroCrossNet reached a classification accuracy of 91.30%, surpassing several strong benchmark methods, including CNN-LSTM, EEGNet, DeepConvNet, spectral CNN, and graph-based EEG decoding frameworks. Furthermore, a large number of ablation studies reveal that the integration of temporally, spectrally, and spatially complementary representations considerably boosts robustness and inter-subject consistency. Full article

For the challenging scenario in which only normal node labels are available in semi-supervised graph anomaly detection, existing generative methods usually synthesize abnormal nodes through random perturbation or feature interpolation. However, these methods fail to consider node abnormality comprehensively from both structural and attribute perspectives, resulting in generated pseudo-anomalies of limited quality and insufficient reliability. In order to address this problem, we propose DGAM (dual-guided anomaly mining), a framework for selecting pseudo-anomaly nodes based on the dual-index measurement of topological anomaly and feature consistency. The core of the framework is the joint anomaly evaluation module, which quantifies node anomaly through two computable metrics. The topological boundary score (TBS) measures the boundary of a node’s topological position based on the proportion of connections between a node and labeled normal nodes in its K-hop neighborhood. The feature deviation score (FDS) evaluates the consistency of a node’s local features by calculating the average cosine similarity between its features and those of its K-hop neighbors. The module selects a fixed set of nodes with higher comprehensive anomaly scores from the labeled normal nodes as pseudo-anomalies, so as to construct a training set containing explicit supervision signals. The model adopts a shared encoder architecture and jointly optimizes the classification loss based on pseudo-labels and the embedding regularization loss of the graph nodes to learn a more discriminative node representation. Experimental results on multiple real-world graph datasets show that DGAM can stably improve anomaly detection performance, effectively verifying the effectiveness of the proposed screening mechanism and joint training strategy. Full article

In industrial quality control, the precise identification of severe structural weld defects is paramount. Traditional 2D image-based detection methods are susceptible to illumination and texture interference, while high-precision 3D laser scanning solutions are costly and impractical for large-scale deployment. To achieve reliable geometric defect detection at low cost, this paper proposes a detection framework based on monocular depth estimation and a dual-attention point cloud network. First, YOLOv8 is employed for rapid region of interest extraction, and an advanced monocular depth estimation model generates 3D pseudo-point clouds containing geometric information. Secondly, addressing the challenge of distinct spatial orientation features in missed weld defects that are prone to confusion, this paper introduces a dual-attention-enhanced point cloud classification network named DA-PointNet++. This model embeds dual-attention modules within the PointNet++ backbone network, enhancing key feature representation in both the channel and spatial dimensions. Experimental results demonstrate that this approach achieves an accuracy of 93.67% and a recall rate of 90.51% in a unified binary classification task for general weld defect detection, effectively identifying both normal welds and complex missed weld defects. Compared to PointConv, Dynamic Graph Convolutional Neural Network (DGCNN), and mainstream Point Cloud Transformer, this method significantly reduces false negative rates while maintaining low computational costs, offering a cost-effective solution for industrial automation. Full article

Drug-induced QT interval prolongation is a critical safety concern in drug development, yet accurate and mechanistically interpretable prediction from chemical structure remains challenging due to the limited substructural resolution of existing approaches. Here, we present a motif-level graph learning framework for interpretable QT risk prediction. In this framework, molecules are decomposed into chemically meaningful motifs, enabling representation at an intermediate structural scale between atoms and predefined structural alerts. Motif features are encoded using a pre-trained chemical language model, and inter-motif relationships are modeled via attention-based graph learning with cross-scale integration. The model is trained and evaluated on two clinically grounded datasets derived from regulatory drug labeling (DIQTA) and real-world pharmacovigilance data (FAERS), achieving strong and consistent predictive performance with robust generalization across data sources. Importantly, motif-level attention reveals that QT liability is associated with the cooperative organization of compact cationic centers and heteroatom-rich, conformationally adaptable scaffolds, rather than isolated functional groups. These patterns are consistent with known determinants of human ether-à-go-go-related (hERG) channel blockade while providing a more structured and chemically specific interpretation beyond conventional structural alerts. Overall, this work establishes a generalizable and interpretable framework for QT risk prediction and highlights motif-level graph learning as an effective strategy for structure-based modeling of adverse drug reactions. Full article

Real-world networks often exhibit strong transitivity with nontrivial local clustering spectra and degree correlations. Such features are not easily modeled in tractable network models, creating an obstacle to the theoretical understanding of such complex network structures. Here, we address this problem using a model for strongly clustered random graphs in which each triad of a random network backbone is closed with a certain probability. Despite the intricate loopy local structure of the graphs obtained, we provide exact expressions for the local clustering spectrum and the degree correlations, filling the gap in the theoretical description of this model for random graphs. In particular, we find positive degree assortativity accompanying high transitivity, and nontrivial structure in the clustering spectrum. Exact asymptotic analytical results, obtained for uncorrelated locally tree-like backbones, are complemented with extensive numerical characterization of finite-size effects. Full article

Objectives: The rise in the frequency of mental health concerns (depression and suicide) expressed on social media calls for reliable, explainable, and efficient computational methods for mental health surveillance. In this paper, we propose an interpretable framework for text-based detection of post- and community-level mental health risk on social media. Methods: The framework combines (i) Secretary Bird Optimization (SBO) for feature selection of informative linguistic and psychological features, (ii) a BERT (Bidirectional Encoder Representations from Transformers)—CNN (Convolutional Neural Network) model for post-level reasoning, and (iii) a Graph Neural Network (GraphSAGE) for community-level reasoning. The graph is estimated based on semantic similarity between posts and author relations, instead of social interactions (e.g., mentions, replies) between authors. We use SHAP and LIME for model interpretability, uncertainty, and calibration analysis to evaluate the trustworthiness of predictions. Results: The model delivers 93.1% accuracy, 0.91 F1-score, and 0.944 ROC-AUC on the eRisk and CLPsych datasets using a strict user-disjoint validation strategy. SBO lowers the number of features by about 38%, leading to better generalization. The graph-based model enables improved learning of post and user representations by capturing relational dependencies. Conclusions: Our approach offers an explainable and robust means of detecting mental health risk from text. Graph-based representations of semantic and authorship interactions enable community-level analyses, while interpretability and uncertainty estimation facilitate possible human-in-the-loop decision-making. This research does not explicitly consider a human-in-the-loop experiment. Full article