Showing 41 papers for 2026-03-17
This paper critically examines GT-BEHRT, highlighting translational gaps between graph-transformer research and clinical deployment in longitudinal EHR prediction. It argues that treating a visit as an unordered bag of codes misses intra-visit structure, while graph-transformer approaches can better capture visit-level relations and long-term temporal dynamics. The review offers insights and practical recommendations for future development and deployment challenges.
We propose FedTREK-LM, a federated framework for personal knowledge graph completion and recommendations using lightweight LLMs and evolving PKGs. By federating model updates and prompting LLMs with structured PKGs, it enables scalable, privacy-preserving personalized recommendations. The approach leverages Kahneman-Tversky optimization to model user decision biases.
The work introduces supervised and unsupervised hypergraph-based pre-training to boost atrial fibrillation prediction in ESUS patients, addressing small cohorts and high-dimensional features. The pre-training strategies aim to improve predictive performance over standard baselines. Empirical results demonstrate improved AF prediction accuracy.
LUMINA presents a Laplacian-unifying mechanism for interpretable neurodevelopmental analysis via a quad-stream Graph Convolutional Network. By fusing multiple Laplacian perspectives, it captures dynamic ROI relationships in fMRI data and yields more interpretable neurodevelopmental insights. The approach advances explainability in brain-network analysis.
Modality-free Graph In-context Alignment (MF-GIA) enables cross-domain, modality-agnostic graph alignment using in-context learning. It reduces reliance on modality-specific encoders by aligning graphs through prompts and contextual reasoning. The framework improves generalization when raw data is inaccessible or pre-vectorized.
MR-GNF introduces multi-resolution graph neural forecasting on ellipsoidal Earth meshes to enable efficient regional weather prediction. The lightweight, physics-aware model operates on a multi-scale ellipsoidal graph to forecast short-term regional weather with reduced computational cost and improved boundary handling. The approach balances accuracy and efficiency for practical deployment.
We propose Node Role-Guided LLMs for dynamic graph clustering, an interpretable framework that uses node role representations to guide LLM-based clustering and provide semantic explanations for cluster formation and evolution. The approach emphasizes transparency for safety-critical domains. Empirical results show improved clustering quality and interpretability.
Gated Graph Attention Networks are developed to predict the duration of large-scale power outages caused by natural disasters. The model handles high-order spatial dependencies with a gating mechanism and graph attention, improving duration estimation under challenging real-world conditions. The approach supports resilience planning.
ST-ResGAT is an explainable spatio-temporal graph neural network that fuses residual graph-attention encoding with GRU temporal aggregation to forecast pavement deterioration. It is designed for resource-constrained deployment and translates PCI forecasts into ASTM-compliant maintenance actions with explanations.
This work proposes a spatiotemporal graph-based deep learning framework for anomaly detection in oil and gas production networks. By modeling interdependencies and temporal dynamics, it addresses distribution shifts and sparse anomalies to improve detection accuracy. The approach aims at robust monitoring of production systems.
STAG-CN models disease onset in beehive sensor networks using a Spatio-Temporal Apiary Graph Convolutional Network. It builds a dual adjacency graph combining physical co-location and climate sensor correlations and processes multivariate IoT sensor data to predict disease onset across hives.
A Multi-Scale Graph Learning Framework with Temporal Consistency Constraints for Financial Fraud Detection in Transaction Networks under Non-Stationary Conditions
AgentTrace: Causal Graph Tracing for Root Cause Analysis in Deployed Multi-Agent Systems
GNNVerifier: Graph-based Verifier for LLM Task Planning
Lost in Aggregation: On a Fundamental Expressivity Limit of Message-Passing Graph Neural Networks
LLM as Graph Kernel: Rethinking Message Passing on Text-Rich Graphs
PiGRAND: Physics-informed Graph Neural Diffusion for Intelligent Additive Manufacturing
TAS-GNN: A Status-Aware Signed Graph Neural Network for Anomaly Detection in Bitcoin Trust Systems
Suppressing Domain-Specific Hallucination in Construction LLMs: A Knowledge Graph Foundation for GraphRAG and QLoRA on River and Sediment Control Technical Standards
Multi-view Attention Fusion of Heterogeneous Hypergraph with Dynamic Behavioral Profiling for Personalized Learning Resource Recommendation
Graph2Video treats dynamic graphs as video-like sequences and uses video models to capture fine-grained temporal order variations, long-range dependencies, and pair-specific relational dynamics that are hard for existing methods. By modeling temporal evolution as a video, it demonstrates improved performance on link prediction.
GraphVLM introduces a systematic benchmark for evaluating Vision-Language Models on multimodal graphs, enabling structured assessment of their alignment and reasoning over graph-connected multimodal data. It targets real-world domains where multimodal information is organized by relations, such as social networks, recommendations, and scientific discovery.
Masked BRep Autoencoder via Hierarchical Graph Transformer presents a self-supervised framework to learn representations from unlabeled CAD boundary representations (BReps). It uses a masked graph autoencoder to reconstruct masked geometries and attributes and a hierarchical graph transformer to capture multi-scale structure, supporting tasks like part classification, segmentation, and feature recognition.
Skeleton Regression proposes a graph-based regression framework for data lying near a low-dimensional manifold with noise. It constructs a skeleton graph to capture geometry, defines graph-based metrics, and applies graph-aware nonparametric regression with feature transforms to estimate the regression function.
HyReaL introduces a hyper-complex space representation for clustering attributed graphs to counter the over-smoothing problem in GCNs. By operating in a hyper-complex space, it yields more discriminative embeddings and improves clustering performance.
Geometric-Aware Quantization (GAQ) preserves SO(3) symmetry in discrete spaces for equivariant GNNs, addressing computational bottlenecks of high-order representations. It proposes a quantization framework that compresses features while maintaining rotational equivariance, enabling faster and memory-efficient molecular simulations.
The paper proposes a fast method to estimate the condition number of sparse matrices using graph neural networks. It designs feature engineering to achieve O(nnz + n) training and inference complexity, offering a scalable alternative to exact decompositions for conditioning estimates.
KEPo studies a poisoning attack on Graph-based Retrieval-Augmented Generation systems, showing that attackers can inject poisoned information into knowledge graphs to manipulate LLM outputs. It discusses the implications and proposes potential defenses to improve GraphRAG robustness.
Multi-hop Reasoning and Retrieval in Embedding Space investigates leveraging knowledge graphs to augment LLMs for multi-hop reasoning. By retrieving structured knowledge to ground embedding-space reasoning, it aims to reduce hallucinations and outdated information.
JobMatchAI presents an intelligent job matching platform that integrates knowledge-graph-informed skill representations, Transformer embeddings, and explainable reranking. It optimizes utility across skill fit, experience, location, salary, and company preferences, with interpretable recommendations.
Brain-Inspired Graph Multi-Agent Systems for LLM Reasoning introduces a multi-agent framework inspired by global workspace theory to improve LLM reasoning. Multiple agents share a common workspace to coordinate and enhance multi-step reasoning, mitigating accuracy collapse.
OMNIA closes the loop by combining structural and semantic reasoning for knowledge graph completion. It uses a two-stage approach that leverages LLMs to complement graph structural information and improve missing-triple inference.
R3-REC proposes Reasoning-Driven Retrieval-Augmented LLMs for sequential recommendations over multi-granular interest signals. It unifies multi-level user intent reasoning, item semantic extraction, long-short interest polarity mining, similar-user enhancement, and reasoning-based matching.
Iterative Semantic Reasoning from Individual to Group Interests for Generative Recommendation with LLMs presents ISRF, a framework that reasons from explicit individual interests up to implicit group interests to produce generative recommendations.
Open Biomedical Knowledge Graphs at Scale introduces Pathways KG and Clinical Trials KG, two open-source biomedical knowledge graphs, and demonstrates federation and AI agent access with Samyama Graph Database to enable scalable, reproducible research.
Multimodal Connectome Fusion via Cross-Attention for Autism Spectrum Disorder Classification Using Graph Learning proposes a framework that fuses rs-fMRI and structural MRI with cross-attention in a graph-learning setup to classify ASD.
GraphSeek: Next-Generation Graph Analytics with LLMs provides an abstraction for scalable analytics by planning over a Semantic Catalyst instead of directly generating graph queries from natural language, enabling efficient multi-query analytics on industry-scale property graphs.
Low-Complexity and Consistent Graphon Estimation from Multiple Networks proposes a histogram-based estimator with low algorithmic complexity to estimate a common graphon from multiple networks with varying sizes and node sets, achieving high accuracy.
The Reasoning Bottleneck in Graph-RAG analyzes why Graph-RAG systems underperform in multi-hop QA, finding that many questions have the answer in the retrieved context but still suffer from reasoning failures. It proposes SPARQL chain-of-thought prompting to decompose questions into graph-pattern queries.
Mitigating KG Quality Issues introduces C2RAG, a robust multi-hop GraphRAG retrieval framework that uses constraint-checked retrieval and generation to mitigate quality issues in knowledge graphs, reducing retrieval drift and hallucinations.
The paper tackles popularity bias caused by long-tailed user-item interactions in collaborative filtering. Although Graph Neural Networks (GNNs) are effective for CF, their message passing can propagate and amplify this bias. It proposes a post-hoc bias correction method that mitigates popularity bias without re-training, yielding more personalized recommendations in experiments.