Showing 21 papers for 2026-06-19
This work introduces a secondary-structure-aware graph neural network for protein representation. The method builds graphs around secondary-structure motifs (e.g., α-helices and β-sheets) and filters edges by hydrogen-bond energy to emphasize stabilizing interactions, aiming to better reflect folding principles. The resulting representations improve downstream protein tasks.
We propose a hybrid GNN-FEM framework for phase-field fracture simulation. The approach combines graph neural networks with finite element methods to model crack evolution efficiently while preserving a variational physics foundation. This yields generalizable surrogate models with reduced computational cost.
We propose Free-Energy Signatures (FES), a spectral descriptor that treats each layer's attention Laplacian as a Hamiltonian. It extracts thermodynamic potentials and spectral-form factors to diagnose reasoning quality and detect hallucinations in large language models. This expands beyond a few eigenvalues to exploit the full spectrum structure.
SEAGAN is a domain-specific, edge-aware graph attention network for dynamic plant processes. It models multiple interacting physiological processes, such as the A-Ci curve for photosynthesis, enabling more accurate inference of leaf-level responses. The method integrates domain knowledge to handle dynamic plant physiology.
We show that proximity graphs (Delaunay, Gabriel, kNN) improve GNNs for forecasting dust source emissions by capturing spatial relationships. The approach enhances spatiotemporal modeling of dust events and pollution spread. This leads to more accurate emission forecasts.
This study explores graph-based deep learning for map generalization by recasting simplification as node movement and aggregation as link prediction within a unified framework. We benchmark GCN, GAT, and GraphSAGE on multi-scale building datasets and find GraphSAGE offers favorable generalization. This work provides guidance for applying GNNs to map generalization tasks.
We address graph structural entanglement that harms GNNs near class boundaries. We propose Boundary Embedding Shaping with Adaptive Contrastive Learning to improve boundary robustness and stabilize predictions. The method reduces boundary vulnerabilities through boundary-aware embedding and learning.
This work studies modality-imbalanced Federated Graph Learning and proposes a data-synthesis-based solution. It tackles two levels of imbalance: client-level missing modalities and node-level attribute gaps, improving cross-client learning. The proposed approach aims to make MM-FGL more practical under real-world modality constraints.
We propose AI Economist Agent, a model-grounded RAG-based framework for economic scenario analysis using LLMs and knowledge graphs. The agent uses planning, evidence retrieval, and theory-grounded reasoning to produce econometric analyses. The framework aims to align NLP-generated narratives with economic theory and data.
HGCN(O) is a self-tuning toolkit for outcome prediction in event-sequence data. It exposes four GCN architectures (O-GCN, T-GCN, TP-GCN, TE-GCN) across different layer types and combines multiple graph representations with temporal dependencies. Experiments show improved accuracy and stability on balanced and unbalanced data.
We model intransitive dominance in tennis forecasting with a temporal directed-graph GNN. The model captures A>B, B>C, C>A patterns to improve predictive power. It achieves competitive accuracy and Brier score compared to established rating systems like Weighted Elo.
MolGraphBench provides a benchmark suite for evaluating GNN architectures on molecular regression tasks. It standardizes evaluation from molecular graphs derived from SMILES, enabling fair comparisons of property-prediction models. The benchmark helps identify strengths and weaknesses of different GNN designs.
Mask-Morph Graph U-Net introduces a generalisable mesh-based surrogate for crashworthiness under large geometric variation. It uses non-shareable edge-specific aggregation through a mask-morph GNN to better capture nonlinear relationships. The approach aims to accelerate crash simulations while maintaining accuracy.
KG-SoftMAP encodes domain knowledge as a soft knowledge-graph prior for Bayesian network structure learning from sparse discrete data. It maximizes a MAP objective by combining the BDeu score with a logit-form prior, enabling more reliable structure discovery when observations are limited.
We propose a graph neural network surrogate for crashworthiness prediction of vehicle panel components. The surrogate aims to replace expensive finite element simulations with fast predictions for nonlinear impact scenarios. We discuss limitations in capturing detailed 3D effects.
This paper investigates multimodal ablations for precipitation nowcasting with a graph neural network system. It evaluates the impact of radar history, numerical weather prediction, surface observations, satellite channels, stochastic noise, and ensemble losses on forecast accuracy. The results identify which modalities most drive performance.
Implicit Semantic-Aware Communication Based on Hypergraph Reasoning proposes semantic-aware communication using hypergraph reasoning to improve semantic inference at the receiver. The approach represents semantic content as hypergraphs to enable deeper relational reasoning.
Detecting Hallucinations for Large Language Model-based Knowledge Graph Reasoning discusses hallucinations in KG reasoning with LLMs; surveys methods to detect and mitigate hallucinations.
GDGU: A Gradient Difference-based Graph Unlearning Method for Cyberattack Localization in Electric Vehicle Charging Networks proposes a gradient-difference-based graph unlearning technique to localize cyberattacks while enabling data deletion without full retraining.
FineREX: Fine-Tuned NER-RE for Human Smuggling Knowledge Graphs introduces a pipeline for building knowledge graphs from legal documents using a fine-tuned LLM for named entity recognition and relation extraction tailored to human smuggling domain.
This paper introduces FundaPod, a multi-persona agent pod platform that employs a knowledge graph memory to support AI-assisted fundamental investment research. Moving beyond traditional finance AI work focused on trading signals, it enables analysts or AI agents to gather evidence, identify business drivers, compare competing viewpoints, and generate investment memos, with the broader aim of producing investment plans that are transparent, reusable, and verifiable, while contributing to the cumulative development of investment knowledge. The platform coordinates multiple agent personas and a structured memory to organize diverse insights for rigorous fundamental analysis.