ABSTRACT Autonomous driving decision‐making requires a deep semantic understanding of traffic scenes. In this paper, we propose the SGVLM (Semantic Graph Vision‐Language Model) architecture: a vision‐language model that enhances autonomous driving decision‐making through depth‐integrated semantic scene graph fusion. Key objects are represented as nodes (category, state) and spatial‐semantic relations as edges, enriched with pixel‐wise depth estimates from Depth‐Anything‐V2 to capture accurate inter‐object distances. These structured graph features are aggregated via a two‐layer Graph Attention Network and projected into the FastVLM's FastViTHD feature space. A cross‐modal triplet fusion layer then jointly integrates graph embeddings, visual features, and natural‐language queries. Crucially, to ensure computational efficiency without compromising the generalization power of the large‐scale backbone, we employ Low‐Rank Adaptation (LoRA), which significantly reduces the number of trainable parameters and accelerates convergence while maintaining pre‐trained performance. Empirical validation on the DriveLM‐nuScenes benchmark demonstrates that SGVLM_7B achieves relative improvements of 25.9% in BLEU‐4 and 18.6% in ROUGE‐L over the InternVL4Drive‐v2 baseline, and attains 94.56% accuracy on collision‐warning decision tasks in our TTSG‐data safety‐critical scenarios. These results confirm that depth‐integrated semantic scene graph fusion substantially enhances the model's ability to generate actionable driving decisions under complex traffic conditions.

DGFFA: Joint multimodal entity-relation extraction via dual-channel graph fusion and fine-grained alignment

MAGF-CCL: Multi-level attentive graph fusion with cross-modal complementary learning for internal control material weaknesses prediction

Synergistic integration of clinical and multi-omics data for early MCI diagnosis using an attention-based graph fusion network.

The current software systems are becoming complicated, heterogeneous and spread out making the task of code analysis a complicated task. The tools used in the traditional program analysis work independently, the statistical analysis, dynamic analysis, inspection of dependencies, vulnerability scanning, and quality assessment are commonly done separately. The result of this fragmentation is a lack of contextual knowledge, decreased explainability and inability to find root causes of defects or vulnerabilities. In order to overcome such shortcomings, the current paper suggests the creation of Opti-Blend, a visual graph-based modeling system of integrated code analysis. Opti-Blend converts several program representations, such as Abstract Syntax Trees (AST), Control Flow Graphs (CFG), Data Flow Graphs (DFG), Program Dependence Graphs (PDG) and Call Graphs, into a Hybrid Program Graph (HPG). The framework proposes a graph fusion mechanism to be used to combine multi-view representations to a semantic model. A query layer of visualization allows people to explain the issues and investigate them through the graph paths and dependencies. The suggested system can assist in defect detection, vulnerability, and code smell identification as well as dependency risk assessment all in a single visual setting. The experimental validation on open-source repositories proves to be a better detection tool and better traceability than individual tools. Opti-Blend is a contribution to a single, understandable and extendable modeling paradigm of next-generation integrated code intelligence systems.

Human action recognition is a key task in computer vision, with widespread applications in virtual reality, intelligent surveillance, and human-computer interaction. Although deep learning methods have made significant progress in this task, existing methods still face challenges, including difficulties in multimodal data fusion, insufficient robustness in complex environments, and a decrease in accuracy when data is missing or modalities are incomplete. To address these challenges, this paper introduces a novel approach by proposing a human action recognition model based on the Multimodal Adaptive Graph Convolutional Network (MAGNet). The core innovation of this work lies in the integration of adaptive graph convolutions and cross-modal self-attention mechanisms to enhance multimodal data fusion. By dynamically adjusting the contribution of each modality, the proposed method addresses the challenge of incomplete data and improves robustness under real-world conditions, such as missing or noisy modalities. Additionally, the incorporation of the VQ-VAE generative model provides an efficient way to handle missing data and generate anatomically consistent pose features, which sets this approach apart from existing methods. Experimental results show that MAGNet achieves state-of-the-art performance on both the NTU RGB+D and UTD-MHAD datasets. Specifically, on the NTU RGB+D dataset, the model achieves 95.2% and 98.8% accuracy on the XSub and XView protocols, respectively, significantly outperforming existing baseline methods. Furthermore, MAGNet demonstrates strong robustness in multimodal data fusion and complex scene adaptation, effectively handling challenges such as occlusion and lighting variation.

Accurate long-term traffic forecasting is pivotal for resilient intelligent transportation systems (ITS), enabling proactive congestion mitigation, energy optimization, and enhanced urban mobility. However, existing methods struggle to capture the intricate interplay of spatial and temporal dependencies in non-Euclidean road networks. Classical autoregressive approaches fail to model nonlinear dynamics, while deep learning techniques—such as RNN-based graph models, attention-driven Transformers, and state-space architectures—often decouple spatial and temporal learning, rely on computationally expensive mechanisms, and exhibit limited scalability and training instability in long-horizon settings. Although recent advances in spatio-temporal fusion and adaptive graph learning partially address multi-scale interactions, they remain constrained by efficiency and the lack of unified global temporal modeling. To overcome these limitations, we propose HG-GFNO (Hybrid Static–Adaptive Graph Convolutions and Graph Fourier Neural Operator), a unified and parameter-efficient framework that combines hybrid graph convolutions for localized and dynamic spatial encoding with a novel GFNO that extends spectral operators to graph domains for linear-complexity long-range temporal modeling. Extensive experiments on four benchmark datasets (PEMS03, PEMS04, PEMS07, and PEMS08) demonstrate that HG-GFNO consistently outperforms state-of-the-art baselines by up to 10.9% in RMSE and 11.9% in MAE across forecasting horizons, while using fewer parameters and exhibiting superior stability. These results position HG-GFNO as a scalable and practical solution for real-world intelligent transportation systems in smart city environments.

A Novel Method for Vulnerability Detection Based on Fusion and Hyperbolic Neural Network Graphs

Spatiotemporal prediction for groundwater heavy metal contamination using Soft-DTW-based clustering and graph neural network framework.

Scalable multi-view graph clustering via tensorized consensus and individual anchor graph fusion

A Query-Aware Multi-Path Knowledge Graph Fusion Approach for Enhancing Retrieval-Augmented Generation in Large Language Models

Cognitive-inspired knowledge graph fusion for gradient-aligned multimodal sentiment analysis

Saliency-aware multi-resolution graph fusion via self-supervised contrastive learning for robust ultrasound endometrial cancer diagnosis

Attitude-constrained interactive multi-model factor graph fusion for integrated navigation

Abstract Accurate and robust underwater localization remains challenging due to limited absolute positioning availability, asynchronous multi-sensor updates, and non-Gaussian measurement disturbances. This paper presents MEKF-SWFGO, an uncertainty-aware hybrid framework that combines a Manifold Extended Kalman Filter (MEKF) front-end with a Sliding-Window Factor Graph Optimizer (SWFGO) back-end for cooperative INS/DVL/LBL navigation. The front-end enforces attitude consistency on the Lie group SO(3), employs a trapezoidal IMU preintegration for stable propagation, and applies a Huber M-estimator to downweight outliers. The back-end jointly refines recent trajectory segments using numerical Jacobian between-factors and adaptive covariance weighting, while an uncertainty-driven switching mechanism balances real-time filtering and delayed global refinement. Extensive experiments on real AUV data demonstrate that MEKF-SWFGO substantially improves localization accuracy and robustness: total position RMSE is reduced from 33.67 m (CEKF), 21.46 m (MEKF) and 17.05 m (UKF) to 12.10 m.Velocity RMSE falls to 0.44 m/s, representing an 19.1% reduction relative to UKF. Yaw RMSE is improved to 0.74°, a 50.4% reduction over CEKF. Statistical metrics and CDF analyses confirm that MEKF-SWFGO yields tighter, less heavy-tailed error distributions in position, velocity and attitude. These results indicate that the proposed uncertainty-aware manifold and sliding-window fusion is a practical and effective solution for resilient underwater navigation in real operating conditions.

Multi-scale fusion graph convolutional networks

MS-GLFGCN: Multi-stream global-local fusion graph convolutional network for skeleton-based gait recognition

A joint topology-data fusion graph network for robust traffic speed prediction with data anomalism

Hierarchical Cross-Modal Contrastive Attention Transformer with Graph Fusion for Multimodal Stress Detection

This paper presents an intelligent educational decision-making framework that integrates multimodal data fusion with structured knowledge graph reasoning to enhance personalized learning experiences. To address challenges associated with heterogeneous data integration and interpretability, we propose the Cognizant Instructional Field Network (CIFNet), a hybrid neural-symbolic architecture. CIFNet combines symbolic representations of learner states with deep contextual embeddings, supporting dynamic and interpretable decision-making processes in educational environments. It jointly models epistemic progression, pedagogical intents, and instructional dependencies while accounting for uncertainty and sparse feedback. Building on CIFNet, we introduce the Pedagogical Inference Controller (PIC), a meta-cognitive strategic layer designed to refine instructional actions through strategic utility estimation, regret-aware adaptation, uncertainty-weighted exploration, and curriculum alignment. By simulating counterfactual instructional outcomes and prioritizing the reduction of knowledge gaps, PIC aims to promote pedagogically coherent and learner-centered interventions. Experimental evaluations across multiple educational datasets indicate that the proposed framework achieves promising improvements over traditional baselines and several recent deep learning models in predictive accuracy and learning-related metrics. While the results demonstrate the potential of combining symbolic reasoning with neural representation learning for more transparent and adaptive educational decision-making, further studies–particularly in real classroom environments–are needed to fully assess the system’s broader applicability and long-term impact.