graph neural networks

High-fidelity update method for multi-temporal point cloud models with semantic-geometric collaborative constraints

Abstract Local incremental updating of point cloud models remains a core challenge in urban digital governance, limited by two well-documented bottlenecks: inaccurate semantic boundary localization and pronounced seam artifacts in texture stitching across fused regions. To address these limitations, we propose a high-fidelity multi-temporal point cloud updating method integrating semantic, geometric and texture information under a unified semantic-geometric collaborative constraint framework. The core contribution lies in cross-stage collaborative optimization across semantic boundary extraction, point cloud registration and texture fusion. Such improvements stem not from breakthroughs in individual sub-algorithms, but from the integrated pipeline that suppresses cumulative error propagation and delivers steady performance improvements across all processing stages. Specifically, we adapt the classic Cloth Simulation Filtering (CSF) into a semantic-adaptive variant (SA-CSF) as the upstream module, adjusting mechanical constraints with adaptive semantic weighting to distinguish true structural boundaries from pseudo-edges caused by dynamic interference. Built upon these boundary outputs, a spatial topology-aware BiResNet pipeline integrates established position-aware convolution and Graph Neural Network architectures to reconstruct topology in occluded regions and mitigate over-reliance on precise initial poses. For texture harmonization, gradient-domain fusion frameworks are extended into a semantic-weighted multi-scale Poisson fusion mechanism with an enhanced Phong model, achieving joint optimization of illumination correction and fine-grained texture retention. Evaluated on point cloud data from Chuzhou University’s Huifeng Campus, the method achieves 92.5% average class accuracy and 2.1 cm 95% Hausdorff distance (95% HD) for semantic boundary extraction, 91.6% registration success rate with 0.21 m Chamfer distance, and 90.10% texture detail preservation rate with 51.24% texture seam visibility. While not top-ranked on either texture metric, the method achieves the optimal overall trade-off between seam visibility reduction and detail preservation. This balance holds notable practical value for engineering applications: in mixed urban scenarios, single-metric methods (e.g., Graph Cut for detail, Mean Value Coordinates for seams) either compromise overall visual coherence or lose high-frequency details critical for fine-grained management. The unified pipeline also reduces scene-specific parameter tuning effort, lowering operational costs for batch projects. This work provides an effective solution for high-fidelity multi-temporal point cloud updating and supports digital twin city construction and urban digital governance.

A Novel Physics-Informed Graph Convolutional Reduced-Order Model for Fluid Flow on Unstructured Meshes

This study proposes a physics-informed graph convolutional reduced-order model, namely Phys-GCN, for high-fidelity and computationally efficient prediction of steady incompressible flow fields. In Phys-GCN, the incompressible Navier–Stokes equations are embedded into the loss function via residual constraints, such that the spatial feature extraction of graph convolutional networks is integrated with the physics-constrained learning strategy of physics-informed neural networks. This mixed design enables the model to capture complex nonlinear flow features while maintaining a clear level of physical interpretability. Benefiting from the node-edge encoding inherent to graph neural networks, Phys-GCN operates directly on unstructured CFD meshes to learn flow features from graph representations constructed using node attributes and adjacency relationships. In doing so, Phys-GCN dispenses with voxelization or SDF preprocessing and fully preserves the local geometric and topological characteristics of the flow domain. The proposed model is systematically evaluated on steady flows past circular and elliptical cylinders, where the predicted velocity and pressure fields are compared against reference CFD solutions in both interpolation and extrapolation scenarios. Results show that, for all physical quantities, the reconstructed steady flow fields achieve mean relative errors below 5%, exhibiting excellent agreement with the CFD benchmark solutions. After offline training, Phys-GCN achieves inference times that are several orders of magnitude faster than conventional CFD solvers, while maintaining comparable predictive accuracy. These findings demonstrate that Phys-GCN provides an accurate and efficient graph-based and physics-informed surrogate for steady flow-field reconstruction on non-uniform, unstructured meshes, thereby laying a solid foundation for future extensions to more complex three-dimensional and compressible flow configurations.

Dynamic Topological Dependency Reasoning Network (DTTRN)

This paper introduces the Dynamic Topological Dependency Reasoning Network (DTTRN), a novel neural architecture designed for real-time learning and adaptation of complex system topology dependencies. Traditional dependency analysis methods rely on static, pre-defined topologies, limiting their effectiveness in dynamic and evolving systems. DTTRN addresses this limitation by employing a multi-layered, variable topology graph neural network where nodes represent system components and edges represent dependencies. Reinforcement learning is utilized to dynamically adjust edge weights and topology structure based on system behavior, optimizing overall performance and robustness. Furthermore, knowledge graph embedding techniques are integrated to enrich node representations with domain knowledge, enhancing reasoning accuracy. The core claim of this work is to build an architecture capable of real-time learning and adaptation of complex system topology dependency relationships, surpassing the constraints of static dependency graphs. DTTRN achieves this through a synergistic combination of graph neural network adaptability, reinforcement learning optimization, and knowledge graph embedding representation, resulting in real-time, self-adaptive reasoning for complex systems, robust to behavioral changes and benefiting from contextual knowledge for improved precision and generalization.

Dynamic Topology-Dependent Neural Network Optimization

This paper presents a novel approach to neural network optimization that dynamically adapts the network topology based on learned dependencies. The core idea is to leverage reinforcement learning, where two agents collaborate: one adjusts connection weights and the other modifies the network's structure (adding, removing, or restructuring connections). A 'dependency graph' guides the agents' decisions, reflecting the learned information dependencies between neurons. This dynamic adaptation addresses the limitations of traditional methods that assume a fixed network topology, particularly when dealing with complex dependencies and non-Euclidean data. The system aims to achieve more efficient training and improved generalization performance by allowing the network to evolve its structure to better represent the underlying data. The optimization process is driven by minimizing a loss function, and the dependency graph is continuously updated based on the error signal. The key contribution lies in the integration of topology adaptation with reinforcement learning, providing a framework for creating inherently adaptive and robust neural networks. The proposed methodology demonstrates potential for significant improvements in network performance across various domains.

Machine learning-empowered electrochemical nanobiosensors: towards intelligent diagnostics and data analysis.

The central conclusion is that computational sophistication cannot repair absent chemical selectivity or non-independent validation: credible intelligent sensing requires mechanism-grounded signals, information-appropriate models, and validation at the level of the intended deployment claim.

Robust Few-Shot Malware Detection in IoT via ACL on Heterogeneous Behavior Graphs

The proliferation of Internet of Things (IoT) devices has expanded the attack surface for sophisticated, polymorphic malware. Traditional intrusion detection systems (IDSs) and standard graph neural networks (GNNs) struggle with the heterogeneity of IoT traffic, adversarial structural perturbations, and the scarcity of labeled data for zero-day threats. This article introduces HeG-Defend, a novel framework unifying adversarial heterogeneous graph contrastive learning (AHGCL) with meta-learning. We propose a methodology to construct heterogeneous behavior graphs (HBGs) from raw traffic, preserving rich semantic contexts via meta-path topology. To combat label scarcity and adversarial fragility, we use a min–max contrastive training objective, where a learnable graph augmenter generates “hard” adversarial views to force the encoder to learn robust, invariant representations. Furthermore, a prototypical meta-learning module is integrated to enable rapid N-way K-shot adaptation to unseen malware. Experiments on the IoT-23 benchmark demonstrate that HeG-Defend achieves a 99.7% $F1$ -score on known classes and improves few-shot detection on zero-day variants by 14.5% over baselines. Crucially, under glass box PGD topology attacks, our model maintains 91.4% accuracy, significantly outperforming traditional GNNs.

PrecipGNN: Physics-Informed Spatiotemporal Graph Neural Networks for High-Resolution Precipitation Nowcasting at Continental Scale

PrecipGNN: Physics-Informed Spatiotemporal Graph Neural Networks for High-Resolution Precipitation Nowcasting at Continental Scale

Neurostructural Correlates and Impulsive Mechanisms of Exercise Dependence in Middle-Aged and Elderly Adults: Evidence from Gray Matter and Structural Covariance Networks

Initial evidence of neurostructural markers underlying exercise dependence in the middle-aged and elderly adults are provided, demonstrating associations with regional GMV and whole-brain structural covariance networks.

GNMLC-Net: a graph neural inference and multi-layer context-aware fusion network for high-precision detection of avionics thermistor solder joints

A Cascaded Refinement and Context-Gated Semantic Injection mechanism is designed to enhance sensitivity to small targets by refining key feature information and improving the quality of feature representations, which effectively mitigates the semantic-spatial mismatch issue in deep networks while enhancing multi-scale feature representation capabilities.

Convergent pathology, divergent pathways: Structural network dissociation of trait and cognitive disinhibition in cocaine use disorder.

This study provides a dimensional framework for understanding the neurobiological organization of behavioral dysregulation in addiction by demonstrating that distinct dimensions of disinhibition in CUD are associated with partially overlapping but largely separable structural connectome architectures.

Vision transformer and graph neural networks based SuryaNet for yoga pose quality assessment

SuryaNet is introduced, a multi-task model with four technical contributions that establishes a new state-of-the-art in terms of quality assessment, while it also provides users with quality feedback.