Showing 33 papers for 2026-06-16
We conduct a comprehensive comparison of 19 graph layer types in GNNs for driving trajectory prediction. We analyze their ability to model spatial interactions and temporal dynamics to inform architecture choices. The results provide practical guidance on which layers best capture the spatiotemporal dependencies for trajectory forecasting.
This paper surveys semantic reasoning in medicine using knowledge graphs across five domains. It highlights how KGs integrate diseases, drugs, symptoms, and patient data to support decision-making, prediction, and personalized care. The discussion covers applications in clinical decision support, disease and treatment outcome prediction, health recommendations, and the challenges and future directions.
We formalize how LLMs' attention mechanisms distort graph structure when reasoning over text-attributed graphs due to graph linearization. We identify rotary positional embeddings as a key factor causing bandwidth-dependent attention decay, and propose mitigation strategies to reduce distortion.
We introduce SpTGNN, a multi-modal spatio-temporal graph neural network with a mixture of experts to predict soil organic carbon. It represents field measurements as a heterogeneous graph, integrating spectral and temporal modalities to address irregular spatial layouts and data heterogeneity.
We propose TCHG, Tri-Trust Conditioned Heterogeneous Graph Learning, for reliable dynamic trust prediction. By disentangling heterogeneous trust signals into separate evidence channels and conditioning the graph learning on three trust factors, the method better captures complex trust dynamics.
This work guides federated graph recommendations with knowledge encoded in large language models. It addresses non-IID client heterogeneity by leveraging global knowledge from LLMs to align local graph embeddings across clients, improving cross-client recommendations. The approach combines semantic guidance with structural aggregation.
ShipNet is a geometric deep-learning surrogate for real-time ship hydrodynamics. It takes hull geometry and speed as inputs and predicts hull-surface pressures and far-field waves, providing fast approximations to high-fidelity CFD.
DYNA augments frozen LLMs with a dynamic temporal knowledge graph memory where events are nodes and time-stamped edges encode relations. It retrieves relevant memory via random walks and centrality during queries to augment the LLM's responses, reducing catastrophic forgetting on temporal recall tasks by about 7%.
We introduce PAMR, a polarity-aware multi-relational model for predicting signed interactions in biological networks. The model distinguishes polar interactions such as activation and inhibition and improves pharmacological predictions.
AL-GNN presents a privacy-preserving, replay-free continual graph learning framework. It enables continual learning from streaming graph data without backpropagation-based rehearsal, addressing privacy and efficiency concerns. The approach uses analytic learning to update models without storing past data.
Adaptive kNN graph model decouples inference latency from accuracy in large-scale kNN classification. It introduces an adaptive graph construction where k is adjusted to data, balancing retrieval speed and predictive performance.
This paper discusses smoothness errors in dynamics models trained on meshes. It reviews oversmoothing in graph neural networks and proposes approaches, including unitary convolutions, to preserve or control smoothness for physical dynamics.
The authors challenge the traditional spectral vs message-passing GNN dichotomy. They show that, in terms of expressive power, spectral and spatial GNNs are largely equivalent, and advocate broader views and new architectural designs.
We present a frequency-aware epilepsy seizure detector using EEG with separate frequency bands analyzed by a graph-convolutional framework. The method achieves interpretable, neurophysiologically relevant detection by leveraging frequency-domain features.
Dynamic link prediction on temporal signed graphs is addressed with a modular temporal enhancement for GNNs. The approach captures evolving structure and sign-consistent interactions to improve predictive accuracy in time-evolving networks.
The paper advocates a unified taxonomy for post-hoc explanations of knowledge graph completions. It formalizes desirable properties, evaluation protocols, and cross-method comparability to improve reproducibility in explainability for KGC.
LLM-WikiRace introduces a benchmark evaluating planning, reasoning, and real-world knowledge in large language models. Models must efficiently navigate Wikipedia hyperlinks to reach a target page, and the study assesses multiple models to reveal strengths and limitations.
Beyond predefined schemas TRACE-KG enables context-enriched knowledge graph generation. It combines text-driven signals with flexible schema design to produce coherent KGs from long technical documents.
The paper provides a theoretical roadmap for fusing foundation models and knowledge graphs, highlighting the impedance mismatch between continuous models and discrete graphs. It categorizes current neuro-symbolic integration strategies into three tiers and outlines principled directions for integration.
AdaSTORM scales LLM reasoning on dynamic graphs via adaptive spatio-temporal multi-agent collaboration. It uses multi-agent collaboration to share reasoning across nodes and time, enabling scalable dynamic-graph reasoning beyond single-context windows.
Link prediction in knowledge graphs often relies on local neighbor aggregation, missing global structure. We propose Model Graph Inductive Learning (MGIL), a framework that integrates local and global graph information to better capture high-level structural patterns for knowledge graph completion. This leads to more accurate and generalizable link prediction.
Leptomeningeal collateral detection on DSA is improved by vessel-graph neural networks. The framework casts collateral detection as a graph learning problem on high-resolution vessel networks, enabling automatic, fine-grained detection beyond CT angiography and reducing inter-rater variability in collateral scoring.
Provenance-enhanced statements capture claims with attribution in knowledge graphs, but existing models treat provenance as semantically neutral. The authors propose a provenance-aware approach (D) that clarifies how attributed claims relate to factual commitment, to one another, and to reasoning, enabling richer inference over capta rather than just data.
LearnOpt recovers latent cognitive structure of standardized exams from historical questions via an exam knowledge graph and constrained optimization, producing personalized study plans. By applying to NEET questions (2016-2024, n=1496), it uncovers a five-category latent skill distribution and generates targeted study strategies.
MAGE-RAG proposes Multigranular Adaptive Graph Evidence for agentic multimodal RAG in long-document QA. It addresses limitations of fixed Top-k or page-level retrieval by enabling retrieval across multiple granularities (text, tables, images, layouts) to balance evidence coverage, noise, and inference cost while preserving page context for multimodal reasoning.
DeepRoot presents a KG-coordinated multi-agent system that converts historical medical texts into verifiable drug-discovery leads. By standardizing and linking prose with medical taxonomies, the framework enables scalable, verifiable drug-discovery pipelines leveraging LLM agents, retrieval, and reasoning over biomedical texts.
Graphical-Probabilistic Modeling of Generative Flows in LLM-Native Software Systems develops principled, modular representations to document and analyze how LLM-generated artifacts flow through software-like systems, addressing the lack of structure in LLM-native development and enabling design-level reasoning and analysis.
The Initial Exploration Problem (IEP) describes how lay users struggle to begin exploring unfamiliar knowledge graphs due to unknown question space, structure, and starting points. The paper theorizes IEP and discusses challenges and potential approaches to ease initial orientation and exploration.
Boosting KG Foundation Models via Enhanced Negative Sampling argues that random negatives are insufficient for zero-shot KG completion across unseen KGs with different vocabularies. It proposes enhanced negative sampling strategies to improve generalization and retrieval of correct links in zero-shot scenarios.
GMN4AD introduces Graph Matching Network for Alzheimer's Diagnosis with test-time domain adaptation across multi-centered MRI data. It addresses inter-site heterogeneity in structural MRI and improves diagnostic accuracy by adapting at test time to new centers.
Spectral sparsification of Laplacian-constrained Gaussian and Husler-Reiss graphical models studies approximating these Laplacian-structured graphical models with sparser graphs while preserving statistical properties, enabling efficient computation in graph-structured statistics and learning.
EventConnector constructs a temporal event graph to mine relations among social events, capturing localized co-fluctuations and lead-lag dynamics. This enables better understanding, retrieval, and forecasting of temporally related events beyond mere semantic similarity.
KnowML targets improving generalization of ML-NIDS by encoding attack semantics with attack knowledge graphs. It argues that feature spaces should capture attack-relevant semantics, enabling better detection of evolving attack variants with limited labeled data.